Methods & Validation
Form MS-MV

How plans are produced — and checked.

Wrong statistics can misdirect a real study, so this page shows the work: the engine's two passes, the rules it is held to, the primary sources behind every method, and what it will not do.

§1

Two passes, not one

Every plan is drafted by one AI model and then re-examined by a second, more careful pass before you see it. The review pass checks that the analysis matches the design and outcome type, that the power analysis is sensible, that the causal diagram follows standard conventions (confounders must be common causes of both exposure and outcome — mediators and colliders are flagged, not adjusted), that observational designs state a missing-data plan, and that the design, analysis, and oversight read are mutually consistent. Where it finds a real problem, the plan is revised before delivery.

If a response arrives incomplete, the app detects the gap and repairs the missing sections rather than shipping a plan with holes — and if anything still cannot be completed, it says so in a banner instead of hiding it.

§2

Rules the engine is held to

These are standing constraints, enforced in the product — not stylistic preferences:

Confidence intervals must match the estimator. Wilson or Clopper–Pearson intervals only for a single proportion or rate; regression coefficients, odds ratios, and hazard ratios get profile-likelihood or Wald intervals; bootstrapped quantities get percentile or bias-corrected intervals.
Mediation with a binary outcome uses a counterfactual / g-computation framework. Baron–Kenny and product-of-coefficients are not offered there, because the odds ratio is non-collapsible.
Group-level exposures (a clinic, a ward, a census tract) with individual-level outcomes require clustering — a random intercept or cluster-robust standard errors — or an explicit justification for its absence.
Observational designs must state a missing-data plan with the missingness assumption named, and participants are never excluded because an exposure could not be measured without the selection-bias risk being flagged.
Events-per-variable is treated as a floor, not a target (Peduzzi 1996), with the modern sample-size view (Riley, Stat Med 2019; van Smeden) cited where it applies.
Table 2 fallacy warnings accompany adjusted-model outputs and forest plots: only the exposure row estimates the effect of interest (Westreich & Greenland 2013).

§3

Validation

The power and sample-size engine uses the exact noncentral t distribution for comparisons of means — not a normal approximation — and its results are validated against R's power.t.test. Rank-based alternatives are sized with the standard asymptotic relative efficiency inflation, and survival calculations size on events via the log-rank method (Schoenfeld 1983).

Every calculator and recommended method carries its primary source. A selection of the library:

t-test / Welch correction	Student, Biometrika 1908;6:1–25 · Welch, Biometrika 1947;34:28–35
Chi-square / continuity correction / exact test	Pearson 1900 · Yates 1934 · Fisher, J R Stat Soc 1922;85:87–94
Single-proportion intervals	Wilson, J Am Stat Assoc 1927;22:209–12
Rank-based tests	Mann & Whitney 1947 · Wilcoxon 1945 · Kruskal & Wallis 1952 · Spearman 1904
Survival analysis	Kaplan & Meier, J Am Stat Assoc 1958;53:457–81 · Mantel 1966 · Schoenfeld, Biometrics 1983;39:499–503
Agreement & reliability	Cohen 1960 · Landis & Koch 1977 · Bland & Altman, Lancet 1986;327:307–10
Sample-size rules for multivariable models	Peduzzi, J Clin Epidemiol 1996;49:1373–9 · Riley, Stat Med 2019 · van Smeden
Effect sizes & power	Cohen, Statistical Power Analysis, 2nd ed. 1988 · Lehr, Stat Med 1992;11:1099–102

§4

Privacy in the pipeline

Uploaded documents are scrubbed of protected health information before the text ever enters the app's state — identifiers are removed on upload and again at generation, and PHI warnings report categories and counts only, never the matched values. What you type is sent to the AI engine to produce your plan; it is not used to train models.

§5

What this tool will not do

MedStatica is a planning aid. It does not replace a statistician, your IRB, or statistical review — and it says so on the product, in the exports, and in the suggested disclosure text every plan ships with. Every plan is an editable, AI-generated starting draft: treat it as a well-informed starting point to verify, not a final answer. Where the honest response is “bring this to a statistician,” the plan is designed to say exactly that.

Built and maintained by a doctorally-trained epidemiologist and biostatistician (DrPH, MPH) with more than a decade in clinical and health-services research. The expert services on the pricing page are performed personally by the same statistician.

← Back to MedStatica · Plans & pricing