Refinement Statistics and R Factors
https://xraytools.com/rfactors
The table of crystal data at the end of a paper, read one number at a time: what each quantity is a statement about, which of them are not comparable with each other, and what the set as a whole cannot tell you. Type in the numbers a paper printed and leave the rest empty.
- Before this
- These are the numbers a refinement reports about itself: a model was adjusted until its calculated amplitudes matched the measured ones, and each figure describes some part of how well that went. Where the calculated ones come from.
- You supply
- Any subset of the numbers a structure determination reports — wavelength and the highest θ measured, completeness, Rint, R1, wR2, the goodness of fit, the reflection, parameter and restraint counts, the difference-map extremes, and the Flack parameter with its standard uncertainty. Every box is optional and nothing is assumed for an empty one.
- Reading it
- This page does not grade a structure. It prints the arithmetic that follows from the numbers — the resolution from Bragg’s law, reflections per parameter, wR2 / R1 — and says what each quantity measures. It quotes no acceptable range for any of them, because what counts as good depends on the compound, the data and the discipline, not on a number this site could check.
Worked examples: a Flack parameter that decides nothing · data that stop early · a refinement held together by restraints
Earlier on the path: Displacement Parameters and NPD Atoms On How to read a published structure, step 4 of 4
Notation here: d · θ, 2θ · F, |F| · R1, wR2, S · Rint · x, u(x) — what each one means here
Terms here: reflection
See also: Friedif Calculator · Difference Map · CIF Parser · Displacement Parameters and NPD Atoms
What each input changes
- highest θ measured
- The highest angle measured, which sets dmin. It bounds the detail in the map and the number of parameters the data can support — not the precision of any one coordinate.
- Absolute structure
- The uncertainty on the Flack parameter, and the whole of whether that parameter means anything. Change it and the page’s reading of the same x changes completely.
Teaching with this page
- Objective
- After this page a learner can read a table of crystal data one number at a time, and say what the set still cannot establish about the chemistry.
- Start from
- this worked example
- Ask first
- A structure refines to R1 = 0.028. Is the chemistry right?
- Watch for
- “Almost certainly — that is an excellent fit”
Check yourself: A structure refines to R1 = 0.028. Is the chemistry right?
The fit says nothing about it either way Almost certainly — that is an excellent fit
An R factor measures how well calculated amplitudes match measured ones. A model with the wrong element at a site, a mis-assigned hydrogen or a chemically impossible molecule can match beautifully, because neighbouring elements scatter almost identically. The chemistry is judged from distances and angles and displacement parameters, not from this table.
Input
These numbers are illustrative
Nothing was submitted, so the boxes hold a made-up set that hangs together — a Mo Kα data set, an anisotropic model, no restraints. They are not a real structure and no compound is being described. Paste your own numbers over them.
What was measured
| Resolution dmin | 0.7696 Å |
|---|---|
| sin θ / λ | 0.6497 Å−1 |
| Completeness | 99.6 % |
| Rint | 0.0412 |
The highest angle measured sets the finest detail the data can carry: dmin = λ / (2 sin θmax), which is Bragg’s law rearranged. It is the finest detail a Fourier sum over those reflections can put into the electron density — a limit on the MAP. It is not a minimum distance between atoms. A refined coordinate is routinely located to a small fraction of dmin, because the refinement adds atom shapes and chemical knowledge the raw Fourier sum does not have. What data stopping early costs is detail in the map, and with it the number of parameters the refinement can support — not the precision of any one of them, and not the R factor, which can look excellent on coarse data. Where 0.6 Å−1 comes from: that is the sin θ / λ a full data set is expected to reach for publication — d = 0.83 Å, which is θ = 25.24° with Mo Kα and 67.68° with Cu Kα, which is why a copper diffractometer has to swing so much further for the same data.
Completeness is what fraction of the reflections that COULD have been measured to that angle actually were. It is the other half of the resolution statement: data taken to a fine d but missing a wedge of reciprocal space carry systematic gaps rather than a uniform loss of detail.
Rint compares reflections that symmetry says must be equal, and it is computed from those equivalents as they are merged: once they have been averaged into one value the differences it measures are no longer there to compare. It is therefore a statement about the MEASUREMENT — counting statistics, absorption, crystal quality, how well the correction worked — and not about the model, which has not been built yet when it is computed. Read it beside the R factor rather than as a bound on it: the two are computed over different observations, on different quantities, with different weights, so neither is a floor or a ceiling for the other.
Three more figures come from the data-processing report rather than from the CIF, and they are about the measurement itself. Mean I/σ(I) in the outermost resolution shell says how far above the noise the weakest, highest-angle reflections were measured; data have conventionally been cut where it falls to about 2, so it is a statement about how much of the stated resolution carries signal. Redundancy, or multiplicity of observation, is the average number of times each unique reflection was measured: a statement about how much averaging stands behind each merged value, and one that Rint depends on as well as on the quality of the data (Diederichs & Karplus 1997). CC1/2 is the correlation between the mean intensities of two random halves of the measurements of each reflection, near 1 at low resolution and falling towards 0.1 at the edge of useful data (Karplus & Diederichs 2012): a statement about whether a shell still holds signal that two independent halves of the experiment agree on.
How well the model fits
| R1 | 0.0389 |
|---|---|
| wR2 | 0.0951 |
| wR2 / R1 | 2.44 |
| Goodness of fit S | 1.043 |
R1 and wR2 are not the same quantity measured two ways. R1 is computed on F, unweighted, and quoted for the reflections a paper calls observed, usually I > 2σ(I). wR2 is computed on F2, weighted, over ALL data including the weak reflections that carry the least information and the largest relative error. Squaring the quantity and keeping the weak data both push it up, which is why the two are quoted together and why wR2 is the larger. The definitions: R1 = Σ‖Fo| − |Fc‖ / Σ|Fo|; wR2 = ; S = . The weight w is the refinement’s own, and SHELXL’s usual form is 1 / [σ²(Fo2) + (aP)² + bP] with P = [max(Fo2, 0) + 2Fc2] / 3 — the scheme a CIF prints in _refine_ls_weighting_details, and the one the CIF report reads back.
The goodness of fit S compares the residuals against the uncertainties the weighting scheme assigned them, over the n − p degrees of freedom. Its expected value is 1 when those uncertainties are right, so a value far from 1 is first of all a statement about the WEIGHTS rather than about the structure — which is why refinement programs that optimise the weighting scheme drive it towards 1 without any atom moving.
What stands behind it
| Reflections per parameter | 15.9 |
|---|---|
| Restraints | 0 |
Every refined parameter has to be paid for with observations. The ratio of independent reflections to parameters is how much data stands behind each number in the model, and it is the reason a structure refined anisotropically needs data a coarser one does not: six displacement parameters per atom instead of one. Restraints are counted separately here because they are not observations — they are chemical knowledge added to the refinement, which is a different kind of support and is worth seeing separately. What counts as enough: checkCIF asks for at least 10 reflections per parameter (PLAT088), relaxed to 8 for a non-centrosymmetric structure with no atom as heavy as argon (PLAT089), where the Friedel pairs are merged because they carry no measurable difference. A ratio below those is a question to answer rather than a fault. A paper’s “data / restraints / parameters” line is three counts, and the ratio is its first over its third, as here. Restraints join the observations in one place only: SHELXL’s restrained goodness of fit, whose denominator is n − p + r for r restraints.
What is left over
| Largest peak | 0.31 e Å−3 |
|---|---|
| Deepest hole | −0.28 e Å−3 |
The difference map is what the model does not account for. Its extremes are quoted as a peak and a hole, in electrons per cubic ångström, and WHERE they sit decides what they mean: features at bonding distances from a heavy atom are usually series termination and absorption, a peak at a chemically sensible position is a missing atom or a disorder component, and a peak on top of an atom is a scattering power the model has got wrong. A number alone cannot tell those apart, which is why this page prints it and does not grade it. The site’s difference map page shows what each of those looks like.
What a checkCIF alert means
checkCIF, the IUCr’s validation service built on PLATON, reads a CIF and lists what it finds as alerts, each with a level:
- Level A — a potentially serious problem with the data has been detected.
- Level B — a potential problem has been detected, and usually some corrective action is required.
- Level C — an outlier to the defined standards has been detected, which may need to be corrected.
- Level G — a general issue to check, not necessarily an error.
An alert is a question to answer, not a verdict. It says a number is outside what the test expects; whether that is an error, an unusual structure or a deliberate choice is for the author to decide and explain, and a set of small alerts can together point to something none of them shows alone. Some a chemist meets often, with what each one tests:
| Alert | Heading | What it tests | Explained on |
|---|---|---|---|
PLAT029 |
Low Value of _diffrn_measured_fraction_theta_full | completeness to the angle at which the data should be complete | completeness, on this page |
PLAT082 |
A High R1 Value is Reported | the agreement factor R1 | R factors, on this page |
PLAT088 |
Poor Data / Parameter Ratio | how many reflections stand behind each refined parameter | reflections per parameter, on this page |
PLAT097 |
Large Reported Maximum (Positive) Residual Density | the highest peak left in the difference map | difference maps |
PLAT220 |
Unusual Ueq(max)/Ueq(Min) range for non-H Atoms in Non-solvent | how widely the equivalent isotropic displacement parameters spread | displacement parameters |
PLAT250 |
Large U3/U1 Ratio for Average U(i,j) Tensor | how elongated the ellipsoids are on average | displacement parameters |
PLAT340 |
Low Bond Precision for C-C in Light Atom Structures (Z(max) < 20) | how precisely the C–C bond lengths are known | distances and angles |
PLAT420 |
D-H Without Acceptor | a hydrogen on a donor atom with no acceptor near it | distances and angles |
PLAT032 |
Standard Uncertainty in Flack Parameter too High | whether u(x) can settle the absolute structure | the Friedif calculator |
What these numbers do not say
These numbers describe how well a model accounts for a measurement. None of them says the chemistry is right: a structure with an excellent R factor can still have the wrong element at a site, a mis-assigned hydrogen, or a molecule that is chemically impossible. That is what the geometry is for — see distances and angles and displacement parameters.
Where this comes from
- Reporting and evaluating absolute-structure and absolute-configuration determinations
H. D. Flack, G. Bernardinelli, J. Appl. Cryst. 2000, 33, 1143–1148 · doi:10.1107/S0021889800007184
The criteria for when a Flack parameter has actually determined an absolute structure, and how the answer differs for a material already known to be enantiopure. This page deliberately quotes no threshold of its own: the condition it depends on is not something a table of numbers can be asked, and a figure this site cannot check is one a reader would quote anyway. - Linking crystallographic model and data quality
P. A. Karplus and K. Diederichs, Science 2012, 336, 1030–1033 · doi:10.1126/science.1218231
Where CC1/2 comes from: the correlation between two random halves of the measurements, proposed as the test of whether a resolution shell still holds signal. - Structure validation in chemical crystallography
Anthony L. Spek, Acta Cryst. D65, 148 (2009) · on the reading list under “The programs a determination actually uses”
The program behind checkCIF, and the paper its alert levels are described in. - checkCIF validation ALERTS: what they mean and how to respond
A. L. Spek, Acta Cryst. E 2020, 76, 1–11 · doi:10.1107/S2056989019016244
What an alert means and how to answer one, for the alerts the panel above lists and the many it does not.