3  Team Science for Biostatisticians

The first email from Dr. X, a cardiologist, is four lines long. The hospital’s data warehouse holds about five thousand heart-failure discharges. Some patients had a home-health visit within a week and some did not. Could you ‘run the numbers’ showing that the visits cut 30-day readmissions, in time for an abstract due next month? As posed, the comparison has a flaw no software will report. A patient readmitted on day 3 never had the chance of a day-5 visit. The sickest patients are therefore counted as unvisited, and home health looks protective whether or not it does anything. The analysis would run, the estimate would be precise, and the paper would be wrong. A conversation held before any code is written prevents it. This chapter follows that conversation from the email to the HF-HOME trial, whose plan, data, and report the rest of the book uses.

The best thing about being a statistician is that you get to play in everyone’s backyard.

John W. Tukey, attributed; quoted in his New York Times obituary (2000)

Diagram of the analysis pipeline: six stages in order (Plan, Acquire, Wrangle, Analyze, Report, Share) above two bands that run under every stage (Version control and Environment). Highlighted for this chapter: Plan.
Figure 3.1: Where this chapter sits: the intake conversation, the scope of work, and the authorship agreement belong to the Plan stage, before any data arrive.
NoteIn this chapter

Reading time. About 35 minutes.

You’ll need. Chapter 2 for the cost-benefit case for a reproducible analysis. The chapter has no code; the trial it introduces is analyzed in Chapter 26 and reported in Chapter 33.

You’ll be able to.

  • Place an engagement on the collaboration spectrum and state the authorship expectation that goes with it.
  • Run an intake conversation that surfaces the red flags in a collaborator’s request, and write the result as a scope of work.
  • Classify a disagreement as preference, best practice, or validity, and give the response each warrants.
  • Decide authorship against the four criteria of the International Committee of Medical Journal Editors (ICMJE), keeping Contributor Roles Taxonomy (CRediT) roles separate, and triage a portfolio of concurrent projects.

Outline.

If these answers come quickly, skip to the next chapter. The answers are in Section 3.12.

  1. What distinguishes a ‘collaborative’ biostatistician role from a ‘consulting’ role?
  2. Name three communication skills essential for biostatisticians on interdisciplinary research teams.
  3. Why does team science succeed or fail on clear expectations about authorship and credit?

3.1 Where you sit: the collaboration spectrum

Most biostatisticians spend the greater part of their careers on teams alongside people who cannot read their code and do not share their vocabulary. Working across that gap is a learned skill, so it can be taught. A statistician described as ‘difficult to work with’ has usually not been rude to anyone. They have failed to establish, in writing and early, what was being asked and what would be delivered. Every later disagreement has then been argued without a shared record of what was agreed. Nearly every collaboration failure in this chapter traces to that missing document.

The first thing to establish is the role. An investigator who says ‘we need a statistician’ may be asking for a methodological consult, a hands-on collaborator, a co-investigator on the grant, or hired labor for the analysis section of a manuscript. The expectations, the timeline, the authorship implications, and the appropriate level of pushback differ for each (Table 3.1).

Table 3.1: The collaboration spectrum. Authority, time, and credit rise together down the table.
Position Mode Your role Authorship expectation Where it is common
Consultant Transactional: a question or dataset arrives, an answer leaves Analysis or methodological advice Usually acknowledgment; the deliverable is the analysis or report Statistical consulting services, industry consulting groups
Collaborator Sustained across design, analysis, and writing Substantively involved in the science Authorship when the ICMJE criteria are met (Section 3.5) Most academic biostatistics
Co-investigator Named on the grant, with allocated effort Partial authority over direction; deeply involved in design Authorship, and effort paid from the grant Mid-career and senior statisticians on funded research
Principal investigator (PI) Leads the project The statistical method is itself the science First or last author Methodological research; less common as lead in clinical research

The common source of tension is mismatch. A statistician may be treated as a consultant while seeing themselves as a collaborator, or the reverse. Dr. X’s email reads like a consulting request (‘run the numbers’). However, the question inside it, whether home health changes outcomes, needs a collaborator from design onward. Naming the position out loud at intake is the first step in avoiding the mismatch.

Question. Dr. X offers to list you in the acknowledgments ‘for the stats’. You expect to design the analysis, write the methods and results, and revise the manuscript. Where on the spectrum is each of you placing the engagement, and when should the difference be raised?

Answer. Dr. X is describing a consultant; the work you expect is a collaborator’s, which carries authorship when the ICMJE criteria are met. Raise it at intake, before any analysis runs, and record the agreement in the scope of work (Section 3.3). Raised at submission, the same conversation is a dispute about work already done.

3.2 The intake conversation

The first meeting with a new collaborator sets the direction of the collaboration. It does so mostly through what is not asked. An investigator arrives with a dataset and a question, and the statistician takes the work. The terms nobody raised (what the deliverable is, when it is due, whether there will be authorship) become assumptions that each party holds privately and differently. They surface months later, usually in a week when everyone is busy.

The remedy is a structured intake, which is less formal than it sounds. It is a fixed set of questions asked in the first conversation and written down afterward (Table 3.2). Sending the questions before the meeting often resolves half the issues without the meeting.

Table 3.2: The intake questions. Write the answers down the same day.
Topic What to ask What the answer decides
The scientific question What are you trying to find out, and what would the answer enable? Push past ‘is A better than B’ to ‘better in what sense, for which patients, in what setting’ The design
The data What exists; how it was collected; what can be shared; whether it exists yet Most of the project’s fate
The deadline A grant deadline in three months, or an exploratory question with none? What is feasible
Authorship Co-author, acknowledgment, or hired help? The position on the spectrum
The deliverables A two-page memo, a Quarto report, a paper, a regulatory submission? The time commitment
Approvals Which institutional review board (IRB) protocol and which data use agreement (DUA) cover this use? Whether the work may start

3.2.1 Red flags

Some answers warrant more questions before you agree to anything:

  • A request to perform a specific test (‘we want a chi-square’) without the hypothesis it is meant to test.
  • A dataset the team has already analyzed several times (‘it didn’t show what we hoped’), which makes any hypothesis now proposed a post hoc one.
  • A comparison defined by what patients received after the starting point, which builds in immortal time.
  • Pressure to produce results by next week.
  • Vague answers about the IRB protocol or the DUA.

Dr. X’s email carries three of these: a fixed answer in search of an analysis (‘showing that the visits cut readmissions’), a deadline, and an exposure defined by what happened after discharge.

3.2.2 Worked example: the intake document

The document below is what the intake conversation with Dr. X produced, for a cohort drawn from the electronic health record (EHR). It reframes the question with a landmark, so that every patient’s exposure is fixed before follow-up begins, and it lists the red flags that remain.

Project. Post-discharge home health visits and 30-day readmission or death in heart-failure patients.

Investigator. Dr. X, Cardiology.

Question. Among heart-failure patients alive and not readmitted 7 days after discharge (the landmark), is receiving at least one home health visit in those 7 days associated with lower all-cause readmission or death over days 8 to 30, after adjustment for measured confounders?

Data. Retrospective EHR cohort, \(n \approx 5{,}000\), obtained from the institutional research data warehouse under existing IRB. Includes demographics, baseline ejection fraction, medication on discharge, date of each home health visit (needed for the landmark), dates of readmission and death.

Deadline. Manuscript draft by month 4; submit month 6.

Methodology (to be confirmed). Propensity-matched cohort with multivariable logistic regression for the primary outcome. Sensitivity analyses with inverse-probability weighting.

Deliverables. Pre-registered analysis plan (month 1), Table 1 and primary results (month 2), manuscript draft (month 4), revisions (month 5).

Authorship. Statistician will be second author. Will be involved in writing methods and results, and revising the full manuscript.

Boundaries. Statistician will not write the introduction or discussion. Statistician will not perform additional analyses requested after month 4 without renegotiation.

Red flags noted. (1) Immortal time: as first posed (‘a visit within 7 days’ against a 30-day outcome), patients readmitted before a visit could happen would count as unexposed, flattering home health. Addressed by the 7-day landmark above. (2) Confounding by indication: referral for home health depends on frailty and social support, which the EHR records poorly; propensity methods adjust only for what is measured, so the result is an association, and a causal answer needs a randomized trial. (3) Outcome capture: admissions elsewhere are missed unless the EHR is linked to the state health-information exchange. (4) Deaths before readmission are counted in the outcome, not treated as non-events. To revisit if EHR data quality is worse than expected at first exploration.

The document removed the immortal time from the question. Its second red flag also told Dr. X what the cohort could not deliver. If the question is whether home health causes fewer readmissions, the answer needs randomization. That decision turned the project into HF-HOME, the randomized trial whose statistical analysis plan (SAP) is the worked example of Chapter 26.

TipVerification

Read your intake document as a stranger would, and confirm that each row of Table 3.2 has an answer in it that a third party could check: a named question, a named data source and approval, dates, an authorship statement, deliverables, and every red flag with what was done about it. A blank or a ‘to be discussed’ is a disagreement scheduled for later.

ImportantLLM prompt

A model is a useful second reader for a request, provided you write your own list first.

Context. I am a biostatistician. Below is an email from a clinical collaborator asking for an analysis. [paste the email, with names, dates, and any data removed]

Constraints. Do not propose an analysis. Do not soften the list for tone. Consider at least: a test named without a hypothesis, a hypothesis formed after looking at the data, exposure defined after the start of follow-up (immortal time), confounding by indication, deadline pressure, missing approvals (IRB, DUA), how the data would be transferred, and the authorship offered against the work requested.

Criteria. Return a table: the quoted phrase, the red flag, why it matters, and the question I should ask. Mark any flag you are unsure of as uncertain.

Failure mode. Models flag the explicit issues (no IRB, no analysis plan, vague hypotheses) and miss the subtle ones: a dataset the team has silently inspected, so that the ‘hypothesis’ is post hoc, and an exposure that can only occur among patients who survived long enough to receive it. Check. Compare the table with the list you wrote before prompting, and with the model answer to exercise 1 below; any flag on neither list goes to a senior colleague before the reply is sent. The email may go to the model only after you have removed patient information (Section 12.3).

Question. Why does the intake document name a 7-day landmark instead of comparing patients with and without ‘a visit within 7 days’?

Answer. Exposure must be fixed before follow-up starts. A patient readmitted or dead on day 3 could not have had a day-5 visit, so the naive comparison counts the sickest patients as unexposed and makes home health look protective. Starting follow-up at day 7, among patients still alive and not readmitted, assigns every patient’s exposure before any outcome is counted. Chapter 26 meets the same trap inside the trial, where an event before the planned visit erases the visit.

3.2.3 Exercises

  1. After HF-HOME reports, Dr. X sends a second email. List every red flag in it, say what each threatens, and draft a reply of no more than six sentences.

    Hi, the trial came back not significant, which was a surprise. The team has been going through the data and the patients who actually got the home visit did much better than usual care. Could you run a quick chi-square on just those patients versus usual care? The abstract deadline is Friday. I’ll email you the spreadsheet from my laptop. Happy to thank you in the acknowledgments!

Phrase Red flag What it threatens
‘the team has been going through the data’ The hypothesis was formed after looking at the outcomes Any p-value from the new comparison is uninterpretable
‘the patients who actually got the home visit’ An as-treated subset of one randomized arm Randomization: patients who received the visit differ from those who did not, and patients with early events could not receive it (immortal time)
‘a quick chi-square’ A test named without a hypothesis or estimand The design dictating nothing and the test dictating everything
‘deadline is Friday’ Deadline pressure Time for a correct analysis or a considered refusal
‘email you the spreadsheet from my laptop’ Trial data held and transferred outside the approved system The data use and security terms of the protocol
‘thank you in the acknowledgments’ Authorship offered without reference to the work ICMJE eligibility, which depends on the work, not the offer

A reply that holds the line: ‘Thanks; I know the result was disappointing. The comparison of visited patients with usual care is not one I can run as a test of home health, because the patients who received a visit differ from those who did not in ways randomization no longer balances, including that patients with early readmissions never had the chance of a visit. The trial’s answer is the as-randomized comparison, and I can help write it up as an informative inconclusive result by Friday. If the team wants the visited-patient analysis, I can run it as a labeled exploratory analysis with appropriate adjustment, after the abstract. Please do not send trial data by email; I will pull it from the trial’s analysis environment.’ The first flag is the one most readers miss, because the email presents it as observation rather than as analysis.

3.2.4 Consulting practice by performance

Reading a red-flag list is not the same skill as producing one in real time, in a room, while a collaborator is talking. The intake conversation is a performance. Like any performance, it improves with rehearsal against a script you have not seen before. The role-play below supplies the client’s half from a realistic transcript and asks you to write your own half from scratch. Do it against a clock (fifteen minutes). Resist the urge to revise the client’s words into a cleaner version of the problem than the one a client actually hands you.

  1. Client-intake role-play. A program manager at a community health system calls about HF-HOME’s home-health arm. There is no script to follow; the call below is close to a verbatim transcript of real first calls.

    ‘Hi, thanks for taking the call. So we’ve been doing these home visits for about a year now as part of the HF-HOME thing, and leadership wants to know if it’s “working.” I guess the real question is whether we should keep funding the nurses who do the visits, because the budget conversation is next month and I need something to bring. We have all the visit data in the system the nurses use to log their stops, and I think the hospital can pull readmissions for us, though I’m not 100% sure what they can share. Ideally I’d love something I can put in a one-pager, like a percentage, “home visits reduced readmissions by X%,” something leadership can act on. Oh, and some of the patients only got one visit and some got like six, does that matter? Also I should say, we changed the visit protocol about four months ago, the nurses used to just check meds and now they also do a home safety walkthrough, so I don’t know if that’s the same program or a different one at this point. Can you have something for us by the 15th?’

    Write (a) the clarifying questions you would ask, in the order you would ask them; (b) the program manager’s request restated as a single analyzable question, naming the population, the exposure, the outcome, and the time window; (c) every point in the call where a data question is actually a study-design decision in disguise, and what decision each one is.

(a) Clarifying questions, in order. Start with scope, because it determines whether anything else is worth asking: ‘When you say “working,” do you mean whether the visits caused fewer readmissions, or whether the program is worth its cost either way?’ These are different questions with different answers. Next, the population and the comparison: ‘Is there a group of patients from the same period who did not get visits, or did everyone discharged with heart failure get offered one?’ Without a comparison group, no causal claim about the visits is possible at all. Only a description of what happened to the visited group is possible. Then the protocol change: ‘The safety walkthrough started four months ago; do you want one answer covering the whole year, or do you want the two protocol periods kept separate?’ Then dose: ‘Is “at least one visit” the exposure you care about, or does the number of visits matter to the budget decision?’ Data access and timeline come last, because they gate feasibility rather than the question itself: ‘What exactly can the hospital share, in what form, under what agreement, and by when could I have it in hand given the 15th?’

(b) The restated question. ‘Among heart-failure patients discharged from this system in [a defined period], is having received at least one home-health visit associated with lower 30-day all-cause readmission, compared with patients discharged in the same period who received no visit, after accounting for the measured differences between the two groups (and, if the data permit, analyzed separately before and after the protocol change)?’ This is deliberately narrower than ‘is it working’. It names a population, an exposure defined at a fixed time (discharge, not some point during follow-up), a comparator, an outcome, and a window. It drops ‘% reduction leadership can act on’, because that is a format for the answer, not the question itself.

(c) Design decisions disguised as data questions. Three stand out. First, ‘can the hospital pull readmissions for us’ sounds like a data-access logistics question. It is actually a design decision about outcome ascertainment. If the hospital can only see readmissions to its own facilities, the program will look better than it is, because readmissions elsewhere are invisible, not absent. Second, ‘does it matter that some got one visit and some got six’ sounds like a request for a dose-response analysis. However, answering it requires deciding, before looking at outcomes, whether dose is a study variable at all or a confounder. Patients healthy enough to attend six visits are not comparable to patients who received one, likely for reasons related to their own prognosis. Treating ‘number of visits’ as an exposure to be compared invites the same immortal-time and confounding-by-indication problems as Dr. X’s original request elsewhere in this chapter. Third, ‘does the protocol change make it a different program’ is phrased as a labeling question. It is in fact a design decision about whether one pooled analysis or two period-specific analyses is the right unit of inference. Pooling without checking for an interaction between period and the visit effect would silently average over two different programs and report neither correctly. A strong answer notes that none of these three can be settled with more data alone. Each needs a decision, made before analysis and recorded, about what the study is actually asking.

3.3 The scope of work

For any non-trivial collaboration, a written scope of work (SOW) turns the intake into commitments. It records a shared understanding and has no legal force. When the project drifts, as projects do, the SOW is the document everyone returns to. It is also what you defend when expectations diverge. Verbal agreements about scope, deadlines, authorship, and analytic decisions are soon forgotten. Send an email summary after every meeting, and write the analysis plan before any analysis runs (Habit 5). Table 3.3 is a template, filled in for the HF-HOME trial that grew out of Dr. X’s intake.

Table 3.3: A scope-of-work template, with the HF-HOME entries.
Element What to write HF-HOME example
Inputs What the collaborator provides, by when, in what format Randomization list and case report form (CRF) exports from the trial’s data management system, monthly from first enrollment
Outputs What you produce, by when Locked SAP before the first outcome is ascertained; Table 1 and primary analysis within six weeks of database lock; methods and results sections of the main paper
Methodology The high-level approach, and where it will be pre-specified As-randomized comparison of 30-day readmission or death; full detail in the SAP (Chapter 26)
Revisions How many rounds are included; what counts as a major change Two rounds of manuscript revision; a new outcome, population, or model is a renegotiation
Authorship and acknowledgment As agreed at intake, against the ICMJE criteria Statistician is an author on the main paper and on any secondary paper using the trial’s analysis; contributions recorded in CRediT terms
Boundaries What you will not do No introduction or discussion; no analyses added after database lock without an SAP amendment; no work on data outside the approved environment

The boundaries row is the one most often left blank and most often needed. It is where ‘one more quick analysis’ gets an answer that was written before anyone was annoyed.

Question. Two months after database lock, Dr. X asks for a subgroup analysis by ejection-fraction stratum for a conference. Which rows of the SOW govern the request, and what do they tell you to do?

Answer. Boundaries (no analyses added after lock without an SAP amendment) and Revisions (a new analysis is a renegotiation). Neither says no. They say the analysis is exploratory unless the SAP pre-specified it, that it must be labeled as such, and that the time it takes is a new commitment to be agreed, not absorbed. A subgroup the SAP did name is already in scope.

3.3.1 Exercises

  1. Each request below arrives during HF-HOME. Name the SOW row that governs it and say whether it is in scope.
    1. A reviewer asks for the primary analysis adjusted for site. (b) Dr. X asks you to draft the discussion section.
    2. A site coordinator offers to send you an updated CRF export as an email attachment. (d) Dr. X asks for a second paper on 90-day mortality.
  1. Revisions and Methodology: in scope if the SAP names site as a stratification factor, as HF-HOME’s does, since adjusting for stratification factors is standard; a response to a reviewer is one of the agreed revision rounds. (b) Boundaries: out of scope; you may comment on the discussion, but drafting it is the investigators’ work unless renegotiated. (c) Inputs and Boundaries: out of scope as offered, because inputs come from the trial’s data management system and you do not work on data outside the approved environment; ask for the export through the agreed route. (d) Outputs and Authorship: a new paper with a new outcome is a new scope of work, with its own analysis plan and its own authorship agreement, even though the data are the same.

3.4 Handling disagreement

Not every methodological choice is worth a confrontation. A statistician who cannot tell preference from principle comes across as pedantic. When you and a collaborator disagree, the response depends on what is at stake (Table 3.4).

Table 3.4: Three tiers of disagreement and the response to each.
Tier What it covers HF-HOME example Your response
Preference Coding style, the reference level of a factor, plotting choices, the order of a table Dr. X wants usual care as the second column of Table 1 Defer, unless there is a substantive reason not to
Best practice Multiplicity correction, intention-to-treat analysis, missing-data handling, sensitivity analyses Dr. X wants to drop the tipping-point analysis of unascertained outcomes to save space Argue for the standard; if overruled, record the dissent in the SAP or the methods
Validity An inappropriate test, data dredging, a hypothesis chosen after seeing the data, dichotomizing a continuous outcome, comparisons that break randomization Dr. X wants the visited patients compared with usual care as the trial’s result Push back firmly; if the analysis is still methodologically wrong, withdraw from it

The middle tier is where most disagreements live and where judgment is most consequential. Picking too many fights makes you a difficult collaborator; picking none makes you complicit in bad analyses. Withdrawing, at the third tier, means declining to put your name or your analysis to the result. It should be said plainly and once. Chapter 32 catalogs the cases where it is warranted.

3.4.1 Talking about results

How you report a result to a collaborator is the subject of Chapter 33, which owns the memo, the talk, and the response to reviewers. Three points belong here, because they come up in every collaboration before anything is written.

Report the quantity the listener will act on. That is an effect size with its uncertainty, not a p-value. Consider ‘the treatment reduced HbA1c (glycated hemoglobin) by 0.6 percentage points (95% confidence interval [CI] 0.3 to 0.9)’. It says more than ‘the effect was significant at p < 0.001’, which is the p-value that interval implies. In a meeting, lead with the conclusion and give the support if asked. In a paper, write the conclusion in plain language and include the quantitative support.

Match the verb to the evidence (Habit 11). The three registers in Table 3.5 are different claims, and Chapter 33 and Chapter 32 build on them.

Table 3.5: Three registers of a claim. An inconclusive trial such as HF-HOME supports the second and third, not the first.
Register What it claims When the evidence supports it
‘We found X’ X is established by this study The interval excludes no effect, and the design licenses the verb
‘The data are consistent with X’ X is among the values the data support X lies inside the interval, along with other values
‘We cannot rule out X’ X has not been excluded X lies inside the interval, and so does no effect

Say which kind of interval it is. ‘For 60-year-old patients with baseline HbA1c of 8%, the model estimates an average of about 7.4% under treatment (95% CI 7.1 to 7.7%)’ describes the average patient like this. A single patient’s value is far less certain. With a residual standard deviation of 0.6 points, the 95% prediction interval for one such patient runs from about 6.2 to 8.6%. A clinician who reads the narrower interval as a forecast for the patient in front of them has been misinformed.

WarningPitfall

A disagreement settled by phone leaves no record, and the analysis that ships is the one you objected to. When you are overruled on a best-practice question, the dissent exists only if it is written: an SAP amendment, a sentence in the methods, or at minimum an email summary sent the same day. Without it, the choice reads later as yours.

Question. Dr. X proposes reporting HF-HOME’s primary outcome as ‘home health reduced readmission or death’, citing the point estimate. Which tier is the disagreement, and which register does the result support?

Answer. Validity: the sentence claims an effect the interval does not establish, which misrepresents the result. HF-HOME’s interval includes no effect, so the result supports ‘the data are consistent with a modest reduction’ and ‘we cannot rule out no effect’, and the plain summary is that the trial was inconclusive. Chapter 33 works the memo through with the computed numbers.

3.4.2 Exercises

  1. Classify each request as preference, best practice, or validity, and give your response in one sentence. (a) Report odds ratios rather than risk differences in a secondary table.
    1. Dichotomize N-terminal pro-B-type natriuretic peptide (NT-proBNP) at its median ‘to make the result easier to explain’ in a press release. (c) Omit the adjustment for multiple secondary outcomes. (d) Use a different color palette for the figures.
  1. Preference, provided the primary estimand (a risk difference in HF-HOME) is reported as pre-specified: defer, and suggest the risk difference beside the odds ratio for readers. (b) Validity: dichotomizing a continuous measure discards information, makes the result depend on an arbitrary cut, and here serves a press release; decline, and offer a plain-language statement of the continuous result instead. (c) Best practice: argue for the pre-specified procedure, and if overruled, state in the methods that no adjustment was made and label the secondary results exploratory. (d) Preference: defer, checking only that the palette is legible to color-blind readers.

3.5 Authorship and credit

Authorship allocates credit in an academic system where credit allocates grants, promotions, and standing. Unresolved authorship disputes sour collaborations and sometimes end them. The ICMJE defines authorship by four criteria, all four of which must be met (International Committee of Medical Journal Editors, 2026):

  1. Substantial contributions to the conception or design of the work; or the acquisition, analysis, or interpretation of data.
  2. Drafting the work or revising it critically for important intellectual content.
  3. Final approval of the version to be published.
  4. Agreement to be accountable for all aspects of the work.

The ICMJE also asks that anyone who meets the first criterion be given the opportunity to meet the other three, by taking part in drafting, review, and approval. Contributors who do not qualify, including those whose only role was funding, general supervision, or administrative support, are acknowledged.

CRediT is a separate question. CRediT describes what each contributor did, in fourteen named roles such as Conceptualization, Formal analysis, Data curation, and Writing – review & editing (Brand et al., 2015). It does not decide who is an author; the ICMJE criteria do. A contributor can hold several CRediT roles and still not be an author, and a journal that asks for a CRediT statement is asking for a description, not a ruling. Table 3.6 applies both to the HF-HOME main paper.

Table 3.6: Authorship scenarios for the HF-HOME main paper. ICMJE decides the ‘Decision’ column; CRediT only describes.
Contributor ICMJE 1 2 3 4 Decision CRediT roles
Lead statistician: wrote the SAP, ran the analysis, drafted the methods and results, approved the final version Yes Yes Yes Yes Author Methodology; Formal analysis; Software; Writing – original draft
Co-analyst: independently reprogrammed the primary analysis; has not yet seen the manuscript Yes Not yet Not yet Not yet Offer drafting and approval; author if all four are then met Validation; Software
Site coordinator: enrolled patients and collected the CRF data; reviewed and approved the manuscript Yes Yes Yes Yes Author Investigation; Data curation
Department chair: secured the funding; no other role No No No No Acknowledge Funding acquisition
Statistician who answered one design question by email No No No No Acknowledge None as an author
Statistician who ran a test for an abstract and never saw the abstract Yes No No No Acknowledge, unless offered and accepting a role in drafting and approval Formal analysis

In practice the lead statistician on a clinical trial is routinely an author. A statistician who wrote the analysis but did not see the manuscript before submission is in a weak position, because criteria 3 and 4 cannot be met after the fact. Discussing authorship at the start of the project, in writing, prevents the confrontation at submission.

ImportantLLM prompt

A model can draft a contribution statement; it cannot decide authorship.

Context. Below are the contributors to a manuscript and, for each, a factual list of what they did. The authorship list has already been decided against the ICMJE criteria. [paste names and contributions]

Constraints. Use only the fourteen CRediT role names, spelled as the taxonomy spells them. Do not add, remove, or reorder authors, and do not comment on who should be an author. Assign a role only when a listed contribution supports it.

Criteria. Return a table of author and roles, then the statement in the journal’s format. Flag any author with no supported role.

Failure mode. The model assigns generous roles (Conceptualization for everyone, Writing for anyone who commented) or vague ones (‘contributed to the project’). That is how ghost and gift authorship look on paper. Asked a question about authorship, it answers with CRediT roles as if they settled it. Check. Confirm every role name against the CRediT list, and send each author their own line to confirm before submission; an author who cannot confirm a role does not get it.

Question. A clinical collaborator brings you a dataset and asks for ‘a quick t-test for an abstract’. You run the test and email back the result. The abstract is accepted with you as second author. Is this appropriate?

Answer. Probably not. You meet criterion 1 (analysis of data), but criteria 2 and 3 are unmet: you did not see the abstract before submission, and you cannot be accountable (criterion 4) for text you have not read. Acknowledgment is the more defensible attribution, unless you are given, and take, the chance to revise and approve it. The right time to settle this was the start: ‘I will do the test if I can review the abstract before it goes in’, or ‘this counts as a service contribution; please acknowledge’.

3.5.1 Exercises

  1. For a secondary HF-HOME paper on medication burden, decide authorship for each contributor and give at least one CRediT role: (a) a pharmacist who designed the coding scheme for the discharge medication strings and helped draft the methods, and approved the paper; (b) a data manager who exported the files and has seen no draft; (c) Dr. X, who proposed the question, revised the manuscript, and approved it; (d) a graduate student who checked the medication coding against the source records and is offered, and accepts, a role in revising and approving the paper.
  1. Author: design of a component of the work (criterion 1), drafting (2), approval (3), and, by approving, accountability (4), provided they agree to it; CRediT Methodology, Data curation, Writing – original draft. (b) Acknowledge as things stand: the export alone is a modest contribution to acquisition, and criteria 2 to 4 are unmet; CRediT Data curation describes the role whether or not they become an author. (c) Author: Conceptualization, Writing – review & editing. (d) Author, once all four criteria are met: the verification is a substantial contribution to acquisition and analysis, and the ICMJE asks that such a contributor be given the chance to meet the rest; CRediT Validation, Writing – review & editing. The CRediT roles describe each person in every case; the decision column comes from ICMJE.

3.6 Managing a portfolio of projects

The intake conversation and the scope of work govern a single collaboration, and a collaborative biostatistician never has only one. You will carry five to fifteen active projects at once, at different stages, for different investigators. Each investigator is convinced that their project is the one you are working on today. Survey evidence ranks managing multiple simultaneous projects among the most essential skills in the work, second only to coding for those in their first five years (Slade et al., 2023). It is a skill that no software provides.

The unit of work is the portfolio. A junior statistician experiences the job as a queue: finish project A, then start project B. Projects instead arrive on the investigators’ schedules, block on data and approvals you do not control, and then unblock several at once. What is asked of you is to hold many partly finished projects in a legible state and move each when it is movable. It is not to drive one to completion while the rest stall. Table 3.7 is what a legible portfolio looks like.

Table 3.7: A sample project portfolio, reviewed weekly. The ‘Blocked on’ column names a person, not a state.
Project Investigator Stage Next deliverable Blocked on (owner) Hours this week
HF-HOME main paper Dr. X Analysis Primary results and Table 1 Nothing 12
Readmission risk model Dr. Y Data acquisition Cohort definition memo Data extract (warehouse team, requested on the 3rd) 1
Sepsis registry Dr. Z Revision Response to reviewers Nothing; due in 3 weeks 6
Stroke rehabilitation pilot Dr. W Design Sample-size calculation Choice of primary outcome (Dr. W) 1
Grant resubmission Dr. X Writing Statistical analysis section Nothing; due in 5 weeks 4

Estimate, then double. Statisticians are systematically optimistic about how long an analysis takes. The estimate assumes the clean path, but the work is mostly the mess. There is the merge that reveals duplicated IDs, the outcome coded three ways across sites, the model that will not converge. Estimate the clean path, double it, and add the one review cycle you forgot. Promising month 2 and delivering in month 2 builds more standing than promising month 1 and delivering in month 3.

Make blocking visible, to yourself and to them. Much of a statistician’s calendar is spent waiting: for the data extract, for the IRB amendment, for the collaborator to say which of two outcomes is primary. A blocked project becomes a problem later, usually at the worst moment. For every active project, track the single thing it is waiting on and who owns it. When a PI asks why their analysis has not moved, ‘I am waiting on the corrected data file I requested on the 3rd’ is an answer; silence is not.

Keep the cost of re-entry low. The expensive part of working on many projects is the reconstruction paid each time you return to a project set down three weeks ago. Consider a project that rebuilds with one command (Section 11.1, Chapter 11), whose analytic decisions are fixed in an analysis plan (Chapter 26), and whose history is legible in Git (Chapter 7). It can be put down and picked up in minutes. A project held together in your head and an untracked script cannot. Reproducible infrastructure is usually justified as a courtesy to reviewers and future replicators. Its more immediate benefit is that it lets you carry fifteen projects without dropping any (Habit 1).

Batch the switches you can; budget for the ones you cannot. Context switching is not free. The statistician who answers every email the instant it arrives pays the re-entry cost a hundred times a day. Where the work allows, batch it: a block for one project’s analysis, a separate block for correspondence across all projects. A few switches are forced by genuine urgency and are worth their cost. Most are not, and the skill is telling the two apart.

Capacity is finite, and saying so is part of the job. The boundaries row of the scope of work declines work within a project. The same discipline operates across projects. A statistician who never says ‘I can start this in three weeks, not this week’ ends up doing everything late rather than some things on time. Deferring is not a failure of collaboration. Committing to a deadline you cannot meet is.

None of this requires a particular tool. A single list like Table 3.7, reviewed once a week, is enough.

Question. It is Monday, and four projects are active. A has a grant deadline Friday and is waiting on nothing. B is a manuscript revision due in three weeks, blocked on a corrected data file that has not arrived. C is an exploratory analysis with no deadline whose PI emailed this morning asking for an update. D is a reviewer response due tomorrow and is half drafted. In what order do you work, and what do you send before you start?

Answer. Before starting, send two messages. The first is a reminder to B’s data owner that you remain blocked on the corrected file. The second is a one-line reply to C’s PI with a realistic date, because an unanswered request tends to become a grievance. Then work D first. It is due tomorrow, and its remaining cost is the best known because it is half done. Then A: it is due Friday, unblocked, and the larger unknown, so it deserves the fresh part of the day. B cannot move until the file arrives, so the reminder was your only move on it today. C waits until A and D are clear. The queue model’s error here is working C because its email sits on top of the inbox.

3.6.1 Exercises

  1. You have 20 working hours this week for project work. Your clean-path estimates are: HF-HOME primary results, 8 hours, due Friday; a sepsis-registry reviewer response, 4 hours, due in three weeks; a new intake meeting with Dr. V, 1 hour, requested for this week; a power calculation for Dr. W, 3 hours, blocked until Dr. W chooses the primary outcome. Using ‘estimate, then double’, decide what you commit to this week, what you defer, and what you send to whom.

Doubled, the estimates are 16, 8, 2, and 6 hours. That is 32 hours of work against 20 available, so something must be deferred, and saying so on Monday is the job. Commit to the HF-HOME results (16 hours, due Friday) and the intake meeting (2 hours). A new collaborator’s first impression is set by whether you make time. That leaves 2 hours, not enough for the reviewer response, which is due in three weeks and can start next week. Dr. W’s power calculation cannot start at all; send Dr. W a one-line message that the calculation is waiting on the choice of primary outcome, and offer two candidate outcomes to choose between. Send the sepsis-registry PI a date for the response. The schedule fits because nothing blocked was counted as work, and nothing unblocked was promised that the doubled estimate could not hold.

3.7 Data-management operations

Everything in this chapter so far governs the statistician’s own relationship to a collaborator and a dataset that already exists. A trial the size of HF-HOME also needs someone to decide, before the first patient is enrolled, who is allowed to type data in, who is allowed to look at it, and who is allowed to declare it final. That is data-management operations. On most teams the biostatistician does not own it. They are, however, the person who most needs it to be right, because every downstream number depends on it. Suppose a coordinator has quietly corrected a value in a locked field, or a site has kept enrolling after a stopping rule triggered. Being pulled in to fix that after the fact is a far worse use of your time than naming the rules at intake.

3.7.1 A roles matrix

The operational question is ‘who can do what to which data, and at what stage of the study’ (Table 3.8). A matrix answers it in one place, so that a request (‘can I just fix this one value’) has a documented answer instead of a judgment call made under time pressure.

Table 3.8: A roles matrix for a multi-site trial’s data system. The statistician’s row is deliberately the most restricted: querying and nothing else, so that the analysis can never be accused of having shaped the data it analyzes.
Role Enter new data Query / export Edit after entry Lock fields Unlock a locked record
Site coordinator Yes, own site only Own site only Yes, with an audit trail, before lock No No
Data manager No (unless also a coordinator) All sites Via a documented query resolution only Yes, at scheduled points With PI or sponsor sign-off
Statistician No Yes, read-only extracts No No No
Principal investigator (PI) No Yes No No With documented justification
Monitor or auditor No Yes, read-only No No No

A statistician who can edit source data, even to fix an obvious typo, has compromised the separation between the people who collect a result and the people who evaluate it. The correct channel is a query back to the site, resolved by the coordinator, with both the error and the correction on record. This is the same principle as Habit 2 (raw data are evidence, derived data are a build product), applied to people instead of files. That is, the statistician’s privileged access is to copies and queries, never to the record of what happened.

3.7.2 Participant tracking

A trial’s data system also has to answer a question no single CRF form captures: of everyone who was ever considered for the study, what happened to them. Two logs do most of this work.

The enrollment log records every patient who was approached, in the order they were approached, independent of whether they enrolled. Patient-level enrollment counts by site and week are what a data safety monitoring board asks for first. They cannot be recovered after the fact if the only record of ‘approached but declined’ was in a coordinator’s memory. For HF-HOME, the enrollment log makes it possible to report how many patients were assessed for eligibility before the number who were randomized. That count appears in the consolidated standards of reporting trials (CONSORT) flow diagram. (Chapter 33 covers the diagram itself; this chapter is about the log that feeds it.)

Screen-fail tracking records everyone who was assessed for eligibility and excluded, with the specific criterion that excluded them. This log is the only way to answer a question a reviewer eventually asks. Are the patients who were excluded systematically different from the patients who were randomized, in a way that limits who the trial’s result applies to? Without screen-fail tracking the question cannot be answered, because the excluded patients leave no trace. Table 3.9 sketches what the log needs to hold.

Table 3.9: What a screen-fail log needs, and what each field answers.
Field Why it is tracked
Date and site of screening Detects a site that screens unevenly over time
Eligibility criteria checked, pass or fail on each Lets you compute how many patients failed on each criterion, not just the total
Primary reason for exclusion, if more than one failed Avoids double-counting when ranking exclusion reasons
Consent obtained before screening, yes or no Distinguishes a screen fail from a refusal to be screened at all
Disposition (declined, ineligible, withdrew before randomization) Separates patient choice from study criteria in the CONSORT diagram

3.7.3 REDCap’s role

Most trials of HF-HOME’s size run their data collection in REDCap (Research Electronic Data Capture) (Harris et al., 2009). REDCap’s user-role feature is the usual mechanism for enforcing the matrix above. Each named role (data entry, data manager, read-only) is configured with a specific combination of form-level access, export rights, and the ability to apply or remove a data-entry lock on a completed form. The lock is the operational equivalent of the ‘Lock fields’ column in Table 3.8. Once a form is locked, a role without unlock privilege cannot change it. That is what makes ‘the primary outcome field was locked before any analysis ran’ a verifiable claim about the system, rather than a procedural promise about people’s intentions. REDCap’s Application Programming Interface (API) is a separate layer from this one. This book covers the API as a way to pull data programmatically into a reproducible pipeline. Chapter 14 and Chapter 19 treat REDCap as a data source to be queried. This section treats it as the system that decides who was allowed to put the data there in the first place, and when the data became unchangeable. A statistician setting up a new study’s REDCap project should not configure the roles alone. Decide the matrix with the data manager and the PI first, as part of the scope of work (Table 3.3). Then configure the system to match the decision, rather than the reverse.

WarningPitfall

A locked field unlocked for a convenience edit, and relocked, leaves no visible difference from a field that was never touched. The audit trail (REDCap logs every unlock-edit-relock cycle with a timestamp and user) is the only record. If nobody ever reviews it, the lock controls nothing. Reviewing the audit trail at prespecified points (database lock, before unblinding) belongs in the data-management plan.

Question. A site coordinator emails you, the statistician, asking you to correct a transposed digit in a patient’s discharge date, since ‘it will be faster than going through the data manager.’ What do you do, and which row of Table 3.8 explains why?

Decline, and route the correction back to the coordinator through the documented query process. The statistician’s row has no ‘Edit after entry’ permission for exactly this reason. An edit made by the person who will later analyze the data, however well intentioned, erases the separation between data collection and data evaluation. That separation is what gives the analysis its credibility. The correct sequence is a query to the site, then a source-verified correction by the coordinator with an audit trail. If the field was locked, an unlock authorized under the matrix’s rules comes before the edit and a relock after.

3.7.4 Exercises

  1. For a new two-site trial, draft a roles matrix like Table 3.8 for these five roles: site coordinator, remote data-entry temp hired for one month, unblinded statistician who prepares the data safety monitoring board report, blinded statistician who writes the primary analysis, and the sponsor’s auditor. Name one access right each role must not have, and say what goes wrong if it does.

The remote data-entry temp should have entry rights scoped to a named set of forms and no export or query rights at all. A temporary role with broad export access is the kind of permission that outlives the person holding it, if offboarding is not tied to study milestones. The unblinded statistician preparing the monitoring report needs query access to unblinded data. They must not have access to the forms the blinded statistician’s pipeline reads. Otherwise the separation that protects the primary analysis from knowledge of the interim result collapses. The blinded statistician should have no path to treatment assignment at all, enforced by the system, not by instruction. The auditor needs read-only export with no edit or lock rights. The auditor also needs an audit trail showing what was exported and when, since an auditor’s own access must be auditable.

3.8 Rules of thumb

  1. Habit 5: agree in writing before the analysis runs. The intake document, the scope of work, and the analysis plan are the record every later disagreement is settled against.
  2. Habit 11: match the verb to the design. At intake, ask whether the question needs ‘causes’ or ‘is associated with’, and choose the design that licenses the verb you need.
  3. Habit 1: the compendium is the unit. A project that rebuilds with one command is one you can set down and pick up, which is what makes a portfolio of fifteen possible.

3.9 Cumulative practice

  1. Cumulative (uses Chapter 2). Midway through HF-HOME, Dr. X asks you to skip the analysis plan and the version-controlled compendium ‘just this once’ to make an abstract deadline. Write the reply, using the cost-benefit case of Chapter 2 and the rows of the scope of work that apply.

A strong reply concedes the deadline, refuses the shortcut, and offers what can be done. It names the cost in Dr. X’s terms. An abstract result produced outside the plan cannot be reproduced for the paper, and a number that changes between the abstract and the paper is the event Chapter 2 describes. It cites the SOW’s Outputs and Boundaries rows (a locked SAP before outcomes; no analyses outside the agreed environment). The refusal then rests on an agreement rather than a preference. Finally, it offers an alternative: the pre-specified baseline table, which needs no outcome data, or a submission at the next deadline.

  1. On your own work (no key). List every project you currently have active. For each, write its stage, its single next action, and the one thing it is blocked on, together with who owns that thing. Any project for which you cannot name a next action is the one in trouble; say why.

3.10 Cheat sheet

Table 3.10: Team science at a glance.
Situation Do Tool
A new request arrives Ask the intake questions; name the position on the spectrum Table 3.2, Table 3.1
The request names a test, not a question Ask for the hypothesis and the estimand first Red-flag list
Exposure is defined by what happened after the start Fix exposure before follow-up, with a landmark or randomization Intake document
The intake is done Write the scope of work the same week Table 3.3
A collaborator disagrees Classify the tier before responding Table 3.4
You are overruled on best practice Record the dissent in writing the same day SAP amendment or methods sentence
Reporting a result in conversation Effect size and interval first; match the verb Table 3.5
Deciding authorship ICMJE decides; CRediT describes Table 3.6
Too many projects Double the estimates; name every block and its owner Table 3.7

3.11 Further reading

3.12 Prerequisites answers

  1. Consulting is typically transactional: a question is brought, a statistical analysis is delivered, and the relationship ends. Collaboration is sustained: the statistician is a team member involved across study design, data collection, analysis, and manuscript writing, and shares authorship responsibility for the final product. The middle ground (a few sustained interactions that do not span the whole project) is where most role disputes happen.
  2. Active listening (understanding the scientific question before proposing a method), precise writing (producing analysis plans and memos that non-statisticians can audit), and calibrated language (distinguishing ‘verified’, ‘likely’, ‘assumed’, and ‘unknown’ in reports). Many biostatisticians’ technical skill outpaces their communication; the gap shows up as ‘difficult to work with’.
  3. Authorship determines credit in an academic system where credit allocates grants, promotions, and prestige. Unresolved authorship disputes sour collaborations and sometimes terminate them. ICMJE-style agreement at the outset (and written down) avoids the confrontation at the end.