Biostatistics and Statistical Programming Services

Statistical programming writes cleaning, derivation and output generation once as code, so the analysis can be re-run in minutes when a data query is resolved, late site records arrive or a reviewer requests another sensitivity analysis. Doing this manually takes days and introduces new opportunities for error. Programming also makes the analysis reproducible: a number produced by a script with a logged version can be reproduced years later, whereas a number produced through manual spreadsheet steps usually cannot. Reproducibility is the main reason regulators and journals now request the code.

Independent double programming catches the most errors: a second programmer produces every output from the analysis plan, not from the first programmer’s code, and the two are reconciled before release. Around one output in twenty will disagree, and roughly a third of those errors are in the primary programmer’s work. Other good practices include version control with a change log for every programme, standardised and reviewed macros instead of code copied between studies, data validation and edit checks before analysis, documentation written for whoever inherits the study, and peer review of statistical decisions as well as syntax. Together, these practices make an analysis demonstrably correct.

SDTM organises the collected data in a standard structure that reviewers know how to navigate. ADaM structures the data for analysis, adding the derived variables, flags and populations the statistical analysis plan requires, while preserving traceability so that every value in a table can be traced through ADaM to its SDTM record. Submissions to the FDA, EMA or PMDA need both, together with Define.xml. Journal publications need neither, and building them would add unnecessary cost. The deciding factor is the study’s destination, which we confirm before quoting.

Yes, increasingly, with appropriate care. No major agency mandates SAS. Agencies require that submitted analyses can be reproduced and understood, and R meets this requirement when the environment is documented, with package versions fixed, the session recorded and the analysis delivered as scripts rather than exported results. Several R-based submissions have been accepted. However, reviewers are more familiar with SAS, some sponsors’ standard operating procedures still specify it, and some derived outputs rely on established SAS procedures without an exact R equivalent. For submission studies, we usually recommend SAS for regulated deliverables and R for modelling, simulation and figures, which is also common practice among large sponsors.

We name a missing data strategy in the analysis plan before the extent of missingness is known, because choosing it afterwards can bias the result. Complete-case analysis is common in published work but is the weakest option, as it is unbiased only if data are missing completely at random, which is rarely plausible and seldom tested. Multiple imputation is usually the better primary approach, with an imputation model that includes the outcome and the variables that predict missingness, and mixed models for repeated measures handle some patterns without imputation. The sensitivity analysis is essential: a tipping-point or pattern-mixture analysis shows how far the missing data would need to differ to change the conclusion. We do not use last observation carried forward unless you provide a written justification.

An estimand is a precise statement of what the trial measures. It is now required in protocols because studies were often designed without one. For example, if some participants stop treatment because of side effects, the trial could estimate the effect of being assigned to treatment regardless of what happens afterwards, or the effect of actually taking it. These are different quantities that give different results, and each needs a different analysis. The E9(R1) addendum requires you to state which one you mean before the study starts, by specifying the population, the endpoint, how intercurrent events such as discontinuation and rescue medication are handled, and the population-level summary. This prevents a common and costly disagreement at review.

Yes, and this is a large part of our work. With the data collected, the analysis can still be correctly specified, validated and reported accurately, but it cannot be made pre-specified retrospectively, and an analysis plan written after the results are visible does not carry the same weight. We will still write and date a plan, describe post hoc analyses as post hoc and state what the data can and cannot support. Occasionally this means advising that the study is underpowered for its primary question, which is better identified by us than by a reviewer.

We use the same workflow regardless of dataset size. Every step from raw extract to final output is written as code and run in order from a single entry point, so the whole analysis can be regenerated from scratch. Everything is under version control with a change log, software and package versions are recorded, intermediate datasets are derived rather than edited by hand, and long-running steps are separated so that re-runs do not repeat them unnecessarily. For very large data, such as registries, claims and imaging-derived features, the work moves to a database or a chunked pipeline using the same principles. The test is whether another analyst on a different machine can reproduce your numbers from your data and code.

Programmed figures are generated directly from the analysis. A Kaplan-Meier curve plotted from the ADaM dataset carries the same numbers as the accompanying table, with numbers at risk, censoring marks and confidence bands that help readers judge it. A chart rebuilt by hand can diverge from the table when either is revised, and reviewers notice such mismatches. Programmed figures are also reproducible and easy to restyle, so the same code produces the submission version, the journal version at the required resolution and the conference version. Good visualisation shows uncertainty clearly, and this is decided in the code.

Participant data is handled under HIPAA and GDPR: de-identification is verified, transfer and storage are encrypted, access is limited to your study team, retention is agreed in writing and data are never used to train AI models. Pubrica’s quality management system is certified to ISO 9001:2015. On authorship, ICMJE criteria require substantial contribution to conception or design, or to acquisition, analysis or interpretation; drafting or critically revising the work; approval of the final version; and accountability for it. A statistician who designed the analysis and interpreted the results often meets all four criteria. In that case we tell you and you decide, rather than accepting an acknowledgement for work that merits authorship, which would be a form of ghost authorship. Any AI assistance in programming or drafting is declared.

The calculation itself is straightforward, but the assumptions require careful justification. A two-group calculation needs the effect worth detecting, the variability, the power and the significance level, and you can run it yourself with the free sample size calculator on this site, which shows the formula used. The calculator cannot tell you whether the effect you entered is clinically meaningful, which is the question reviewers ask, and it does not handle clustering, interim analyses, non-inferiority margins, repeated measures or competing risks. Each of these changes the required number substantially and cannot be corrected by inflating the sample size afterwards.

A p-value of 0.06 should be reported as it is, and further analyses will not change its meaning. Report the estimate and its confidence interval so that readers can see what the study supports, which is often a range that includes both meaningful and negligible effects. Do not add covariates until the p-value changes, switch to a one-sided test, remove an outlier that the plan included or describe the result as approaching significance. Statistical reviewers recognise these practices, and they damage the credibility of the work. If the interval is wide, the appropriate conclusion is that the study was too small.

Yes. Funding panels read the sample size section to judge whether applicants understand their own design, so the justification matters more than the number: where the effect size came from, why it is the smallest clinically meaningful difference, what attrition is assumed and why, and what happens if recruitment falls short. We write a defensible section with a calculation that can be reproduced from the information given, and we will tell you early if the design cannot support the aims at a feasible sample size.

Intention-to-treat analysis includes participants in the group to which they were randomised, regardless of what happened afterwards. Per-protocol analysis includes only those who followed the protocol. They answer different questions: intention-to-treat estimates the effect of offering the treatment and preserves randomisation, while per-protocol estimates the effect of receiving it and can be biased because it conditions on post-randomisation behaviour. In a superiority trial, intention-to-treat is the primary analysis and per-protocol is supportive. In a non-inferiority trial the relationship is reversed, because intention-to-treat tends to dilute differences towards the non-inferiority conclusion. ICH E9(R1) provides the framework for specifying which effect you intend to estimate.

Yes, and both must be specified before the data are examined. An unplanned interim analysis inflates the type I error rate and cannot be corrected afterwards. A planned interim analysis uses a spending function, such as O’Brien-Fleming or Pocock, chosen for the required stopping behaviour, so that the overall error rate is preserved and the boundary is known in advance. Blinded sample size re-estimation, in which the nuisance parameter is updated without examining the treatment effect, is the safer option and is often appropriate when the variance assumption was uncertain. Both belong in the protocol and the analysis plan, with dates.