Research Methodology
Most health sites ask you to trust that their article selection is sound. We would rather show you the process. This page documents exactly how a creatine study becomes an article on CreatineIQ — where candidates come from, what disqualifies them, and the weighted rubric used to rank the survivors.
The rubric below is not a description written after the fact. It is read at runtime by the pipeline that produces our research summaries, so the criteria on this page are the criteria that actually score the studies. For who produces the articles and how they are reviewed before publication, see our Editorial Policy.
1. Where candidates come from
Once a week, an automated search queries the National Library of Medicine's PubMed database through its public E-utilities API, covering everything indexed in the previous nine days. The window deliberately overlaps the seven-day schedule so that a paper indexed late is not missed.
The search runs in two tiers. A narrow, high-precision query runs first. Only if nothing survives the eligibility gates does a broader, noisier query run as a fallback. This keeps the typical week's candidate pool small and relevant rather than large and diluted.
2. Hard eligibility gates
Before anything is scored, every candidate must clear a set of pass/fail gates. These are not judgement calls — a paper that fails any one of them is dropped, and the reason is recorded:
- Creatine must be the intervention. A great many papers mention creatine only as the creatine kinase or creatinine biomarker. Those are measurements, not the thing being tested, and they are filtered out.
- The research must be in humans. Animal and in-vitro work is excluded, however interesting the mechanism.
- A substantive abstract must exist. Records with no abstract, or an abstract under 400 characters, are dropped — there is not enough there to summarise without inventing detail. These are reconsidered later rather than permanently excluded, because a thin abstract is usually a temporary state of an online-ahead-of-print record.
- The paper must not already be covered. Every paper we assess is recorded in a coverage ledger keyed by its PubMed ID, with its DOI checked separately so the same study cannot slip through twice under two identifiers.
3. The impact rubric
Surviving candidates are scored on six weighted criteria. Each is an integer from 1 to 5, assessed against published anchor descriptions for every point on the scale. Two rules keep this honest: where an abstract is silent on a criterion — which is very common for funding and conflicts of interest — the uncertainty is scored down rather than assumed away; and no study earns credit merely for being about creatine, since every candidate is.
| Criterion | Weight | What it asks |
|---|---|---|
| Design rigor | 25% | Does the design support the causal claim being made? |
| Sample & power | 15% | How many humans, and was the study sized to detect what it claims? |
| Effect size & clinical significance | 20% | Would the finding change anything for a real person? |
| Novelty | 15% | Does it tell the field something it did not already know? |
| Population reach | 15% | How many of our readers does this actually apply to? |
| Independence (funding & conflicts) | 10% | Who paid, and who benefits? |
What a 5 and a 1 look like
Design rigor
25%Scores 5: Meta-analysis or systematic review of RCTs with a pre-registered protocol, or a large multi-site blinded randomised placebo-controlled trial with a registered trial identifier.
Scores 1: Case report, opinion piece, protocol paper with no results, or any design where the conclusion is not supported by the method.
Sample & power
15%Scores 5: More than 500 participants (or a meta-analysis pooling over 1,000) with an a priori power calculation that was reported and met.
Scores 1: Fewer than 15 participants, or a sample size that is never reported. An unreported sample size scores the minimum rather than being treated as a benign omission.
Effect size & clinical significance
20%Scores 5: A large, precisely estimated effect on a hard clinical or functional outcome, with confidence intervals reported and a magnitude that matters in daily life.
Scores 1: A null result presented as positive, effects appearing only in unplanned subgroups, or an abstract conclusion the reported results do not support.
Novelty
15%Scores 5: First direct evidence on a genuinely open question, a new mechanism or population, or a well-powered result that contradicts the current consensus.
Scores 1: A review that recycles the references of prior reviews, or a result already settled by existing meta-analyses.
Population reach
15%Scores 5: The broad general adult population, or a very large group our readers belong to — older adults, women, vegetarians and vegans, general athletes.
Scores 1: Findings that cannot be generalised to the people reading this site at all: a single case, an extreme environment, or a population defined so narrowly the result is not actionable.
Independence (funding & conflicts)
10%Scores 5: Explicitly funded by a public body, university or independent foundation, with an explicit declaration of no conflicts of interest.
Scores 1: Directly authored, funded or analysed by a company selling the product being tested, with no independent oversight.
4. How the total is calculated
The weighted total is computed in code from the six per-criterion scores — never estimated as a single overall judgement. This matters more than it might sound: it means the ranking is reproducible, and anyone looking at the individual scores can check the arithmetic themselves.
That produces a score between 1.00 and 5.00, which falls into one of four bands:
| Score | Band | What it means for publication |
|---|---|---|
| 4.00 – 5.00 | Strong | Genuinely moves the evidence base. |
| 3.00 – 3.99 | Solid | Worth publishing with the usual caveats. |
| 2.00 – 2.99 | Weak | Publishable only as "here is what came out this week" — the summary must lead with the limitations. |
| 1.00 – 1.99 | Very weak | The reviewer should seriously consider declining to publish it at all. |
A low score does not automatically mean silence. It means the article has to lead with the weaknesses. Several of our published summaries conclude that the study they cover is incremental, underpowered, or that its stated conclusions outrun its own data — which is a more useful thing for a reader to know than a fourth restatement of a settled finding.
5. The coverage ledger
Every paper the pipeline assesses is written to a permanent ledger keyed by PubMed ID, recording the decision and the reason for it — published, or dropped and why. The ledger exists to guarantee no study is ever covered twice, and it doubles as an audit trail of what was considered and rejected, not merely what was published.
Some drops are permanent and some are not. "Creatine is not the intervention" is a permanent judgement. "No abstract yet" is a temporary state, so those records are deliberately left out of the ledger and become eligible again once PubMed catches up.
6. How the calculators are derived
The rubric above governs which research we write about. The dosing calculators are a separate matter, built on published pharmacokinetics rather than on any single study we have summarised. This section is the derivation, the constants, and — because it matters more than the rest — where the model is weakest.
The model
Your muscle holds a pool of creatine with a physiological ceiling. Supplementing adds to that pool; spontaneous conversion to creatinine drains it. Loading is finished when the pool reaches the saturation target. Each day:
The term that matters most is retention. An ingested gram is not a deposited gram: uptake is transporter-mediated and falls off on two independent axes — as stores fill there is less headroom, and as a single dose grows a larger share spills into urine. Even on the first day well under half of a 20 g dose is retained, and the share keeps falling as stores fill: Harris 1992 measured about a third of the dose retained over the first two days at 30 g/day.
baseline_input is not guessed at. Absent supplementation the pool sits stable at its baseline, so endogenous synthesis and dietary creatine together must supply exactly that baseline's daily turnover. Maintenance then only has to cover the extra turnover of a fuller pool — which is why a vegan, whose baseline is lower, needs slightly more than a high-meat eater.
Constants
| Parameter | Value | Basis |
|---|---|---|
| Muscle saturation ceiling | 5.07 g/kg muscle | Roughly 155 mmol/kg dry mass, the widely reported upper limit of muscle total creatine. |
| Skeletal muscle / lean mass | 0.52 male · 0.47 female · 0.49 unspecified | Total skeletal muscle is ~50-55% of fat-free mass in men and ~45-50% in women. Sex is optional; the midpoint is used when it is not given. |
| Daily turnover | 2.0% of the pool per day | Creatine converts to creatinine by spontaneous non-enzymatic cyclization at ~1.6-2.0%/day. A chemical process, so exercise does not accelerate it. |
| Baseline saturation | 0.76 vegan · 0.81 omnivore · 0.84 high-meat | Measured baselines are ~117-125 mmol/kg dm in vegetarians against ~125-130 in omnivores. The real spread is about five points, not twenty. |
| Retention ceiling | 1.87 | Calibrated, not measured. Fitted so the reference subject matches the two literature duration anchors below. |
| Dose half-saturation | 19 g | Calibrated. Sets how sharply retention falls as a single dose grows and more spills into urine. |
| Headroom exponent | 0.5 | Reproduces the roughly-halving of retention observed across a loading phase in Harris 1992. The duration anchors alone do not identify it. |
| Saturation target | 97% of ceiling | "Saturated" in the literature means near-maximal. Uptake approaches the ceiling asymptotically, so a 100% target would never be reached. |
| Training uptake bonus | ×1.15 | Exercise increases muscle creatine uptake (the one-legged exercise arm of Harris 1992). |
Weight-scaled dosing in Standard mode uses the conventional per-kilogram forms of Hultman 1996's fixed 20 g/day loading and 2 g/day maintenance — 0.3 g/kg/day and 0.03 g/kg/day — for Hultman's six loading days, bounded by the safety limits below.
Safety bounds
| Bound | Value | Why |
|---|---|---|
| Maximum daily intake | 25 g | A conservative ceiling rather than the edge of the literature: Harris 1992 loaded at up to 30 g/day (6 × 5 g), and several trials have used 25 g/day. The calculator will not recommend more than this in a day, and refuses rather than extrapolating. |
| Maximum loading dose | 20 g | The conventional loading dose. Weight-scaled ratios are clamped here rather than scaling without bound. |
| Minimum maintenance | 3 g | The floor of the 3-5 g/day range in the ISSN position stand, reported even when modelled turnover is lower. |
| Maximum single serving | 5 g | Larger single doses raise urinary loss and the risk of GI upset, which is why loading protocols split the day into 5 g servings. Two studied protocols (the Alzheimer's pilot and the sleep-deprivation dose) exceed this, and the calculator says so when it shows them. |
Calibration
Two of the constants above — the retention ceiling and the dose half-saturation — are not measured values. They are fitted so that a reference subject (80 kg male at 15% body fat, so 68 kg lean mass, omnivore) reproduces the two best-replicated durations in the literature:
| Protocol | This model | Literature |
|---|---|---|
| 20 g/day | 6 days | 5–7 days (Harris 1992; Hultman 1996) |
| 3 g/day | 28 days | ~28 days (Hultman 1996) |
Both anchors are enforced by automated tests, so any later change to the baselines, the muscle-mass ratio or the saturation ceiling fails the build unless the retention curve is refitted alongside it.
Dose limits, and what they can change
Tailored, Optimal and Brain Health mode each ask for the most creatine you can take in a day. In Tailored and Optimal that number becomes the loading dose and the model works out how long loading takes at it. On the brain-health tier what it can do depends on the shape of the dose, and the difference is the whole reason some protocols are refused rather than reduced:
| Shape of the dose | Where it appears | If your limit is below it |
|---|---|---|
| A rate, loading to saturation | General Brain Health (20 g/day arm); Sarcopenia & Healthy Aging (optional 20 g/day load) | Runs at your limit, and the phase lengthens wherever the model above says your limit needs longer to get there. The endpoint is the same one the studies reach. If at your weight the model says your limit could not saturate within a year, the loading arm is left out and the ongoing dose is still given. |
| A studied range | Menopause & Brain Fog (5–10 g/day); Sarcopenia & Healthy Aging ongoing (3–5 g/day) | Trimmed from the top. The bottom of the range is itself a studied dose, so the protocol stays inside what was trialled. |
| A fixed daily rate | Alzheimer's; MDD; the 5 g/day general arm | Refused, with the studied dose named. We do not scale a trial's dose down to fit a limit it never tested. |
| A single acute dose | Sleep Deprivation | Refused. Where the smaller trialled dose fits your limit, or would fit under the 25 g ceiling, we point you at it instead. |
The refusals are the deliberate part. A loading phase is a rate, so slowing it down still arrives at the same place, just later. A fixed dose is the amount a trial administered for the duration it administered it, and there is no slower route to the same finding — printing a smaller number would be our extrapolation wearing the trial's credibility. Two of these protocols are clinical indications, where that would be a genuinely bad thing to do. So we state the studied dose, say why it cannot be adjusted, and leave the decision with you and your clinician.
Why there is no concussion protocol
Earlier versions offered one, dosed at the 0.4 g/kg/day used by Sakellaris 2006 and 2008. Both papers report the same open-label cohort of 39 children and adolescents with traumatic brain injury serious enough that their outcomes included time on a ventilator and in intensive care. That is not concussion, and it is not adults: at adult body weights the same ratio is 24–36 g/day for six months, a dose nobody has studied. With no adult or mild-TBI trial to take a dose from, there is nothing honest for a calculator to compute, so the protocol was withdrawn rather than relabelled.
Why a single acute dose cannot be spread out
Sleep deprivation is the one protocol here where the intervention is a single dose taken a few hours before the task, rather than a regimen taken daily. Two have been trialled: 0.35 g/kg (Gordji-Nejad 2024) and 0.20 g/kg (Nutrients 2026, the smaller effect). Halving one across two days is not a slower version of the same thing — it is a different intervention that nobody has tested.
Because these scale with body weight, 0.35 g/kg passes 25 g at about 71 kg (157 lb) and 0.20 g/kg at about 125 kg (276 lb). Where your limit rules out the higher dose, the calculator points you at the lower one by name, since it is a real trial dose rather than an interpolation — and if that one does not fit your limit either, it says how far to raise it. Only where the lower dose is also past 25 g at your weight does it decline outright.
Limitations
- Two constants are fitted, not measured. The retention curve is calibrated against two duration anchors. A curve that reproduces both endpoints is not necessarily right about everything between them.
- Calibrated on one reference subject. A 68 kg-lean-mass male omnivore. Accuracy away from that build is an extrapolation.
- Individual variation is not modelled. Baseline stores, transporter density and response to supplementation all vary between people in ways body composition does not capture. Some people are non-responders.
- Lean mass is estimated, not measured. Body fat percentage entered by the user is itself usually an estimate, and error there propagates directly into pool size.
- The maintenance floor overrides the model. For most body sizes the derived requirement is under 3 g/day. We report the ISSN floor instead, and show the derived figure alongside it rather than hiding the difference.
- Brain-health doses are not this model. They are published trial protocols reproduced at your body weight, and each is labelled with the strength of the evidence behind it — which ranges from meta-analyses to a single open-label pilot of twenty people. Several are not established treatments. The one exception is the loading phase. The loading days shown for General Brain Health and for Healthy Aging's optional load are solved with the retention model above at your weight, so those days come from the model even though the doses do not.
- A brain-health loading phase assumes a body composition. That form asks for total weight, not body fat, so solving a loading phase takes lean mass at the reference subject's 15% body fat. At that composition the model returns 6 days for a 20 g/day load, which is what the source studies report, but a reader far from it will load faster or slower than the figure shown.
- A dose limit cannot reshape a fixed dose. Loading is a rate, so a lower daily limit slows it rather than ruling it out. The rest of the brain-health protocols are the amount a trial gave for the duration it gave it, with no slower route to the same place, so a limit below one of them is refused rather than scaled down — a smaller dose would be our extrapolation, not the trial's finding.
The studies each figure is calibrated against are cited in the sections above. The wider literature on creatine dosing and safety is catalogued on The Science page.
Where this can go wrong
A documented process is not an infallible one, and it would be misleading to present it as such. Scoring from abstracts means a paper whose abstract undersells or oversells its own methods can be mis-ranked. A weekly cadence means we cover what was published recently rather than what is most important overall. And a rubric encodes a particular set of priorities — ours weights clinical meaningfulness and independence heavily, which another reasonable person might balance differently.
Publishing the criteria is what makes those trade-offs arguable instead of invisible. If you think we have scored something wrongly, our corrections process is the place to say so.