Skip to main content

Docs · Research

Data: statistics, graphs, Python and R

Every dataset opens in a studio with SPSS-style menus: Data (select, sort, restructure, merge), Transform (recode, standardize, rank, bin, impute…), Analyze (45 analyses from frequencies to Cox regression, factor analysis and clustering), Graphs, a Python/R code runner and an Assistant tab your own AI agent works through. Every result is saved in the Output viewer with the syntax that reproduces it.

Updated 2026-11-09Open a project → Data

The studio at a glance

Project → Data lists the datasets; creating or importing one opens the studio: six tabs (Data · Variables · Output · Graphs · Code · Assistant) under a menu bar (File · Data · Transform · Analyze · Graphs · Code). The Analyze menu opens with a search box and a “Common” shortlist in plain language (compare two groups, predict a yes/no outcome…) above the full catalogue.

TabWhat it holds
DataThe case grid: click a cell and type, Enter moves down, Tab moves right, paste a block from a spreadsheet. Toolbar: undo/redo, add or delete rows, add variable, compute, select cases, pivot, export, history, value labels.
VariablesThe codebook: one row per variable with name, label, type, measure, value labels, missing values and width.
OutputEvery saved result: analyses, code runs and AI answers, pinned ones first. Tables copy as tab-separated text; charts export as SVG or 300-DPI PNG; the syntax that reproduces an analysis is one click away.
GraphsThe chart builder bound to this dataset: bar, line, scatter, box, histogram, Kaplan–Meier, bubble, forest and funnel plots with journal presets.
CodeWrite Python or R against the dataset and run it in the sandbox; the script, printed output, tables and figures are saved together.
AssistantWhere your AI agent works on this dataset: it describes the variables, runs analyses from the catalogue and sandboxed Python or R, and the outputs land in Output. Where an administrator has switched platform AI on, you can also type the question here.

The graphs hub tab moved here: figures belong with the data they draw. The datasets page also has project-wide Outputs and Graphs tabs.

Data files — what comes in and what goes out

Import file (Project → Data) takes CSV, TSV, Excel, SPSS, Stata and R files. All of them are read in your browser: you see the variables, labels and types before anything is stored, and the file itself is never uploaded unless you confirm. SPSS, Stata and R files bring their own dictionaries. File → Export writes the same formats back, instantly, with the dictionary as it stands in the Variables view.

FormatComes in withGoes out with
CSV / TSVColumn headers; types detected from the valuesActive or all cases; value labels in place of codes on request
Excel (.xlsx)The sheet you choose; headers and detected typesData sheet with a frozen header and filters, plus a Variables sheet (the dictionary)
SPSS (.sav, .zsav)Variable labels, value labels, user-missing values (codes and ranges), measurement level, the case weight, datesAll of these, including the weight — the file opens in SPSS with Weight Cases on
Stata (.dta, Stata 13+)Variable labels, value labels, dates and datetimes, strL text; extended missing values (.a–.z) are reported and read as missingFormat 118 (Stata 14+): labels and whole-number value labels; missing codes stay as values (Stata has no missing rules); weighting is a command option in Stata
R (.rds, a data frame)Factor levels as value labels (ordered → ordinal), logicals, Date and POSIXct, haven labels and missing rulesA data frame as saveRDS writes it: factors for coded variables, haven labels where a missing rule exists, Date and POSIXct

Nothing here depends on R or on a server-side converter: the readers and writers are Hikma's own and are checked against SPSS's and Stata's reference reader (ReadStat) and against R itself. Only Stata files older than Stata 13 and R files saved with bzip2 or xz are read on the server, in the isolated sandbox.

Limits: up to 500 variables and the row ceiling shown in the wizard per import; exports write every case up to the same ceiling (use the CSV export for more). The desktop app opens the same files locally and saves them with Save data as….

Data menu — cases

  • Select cases… keeps the cases where a condition is true (age >= 18 AND sex = "F"); the others are greyed out and left out of every analysis unless you untick “Active cases only”.
  • Weight cases… (SPSS Weight Cases) makes a numeric variable a frequency weight: a case with weight 3 counts as three cases in every analysis, table and export, N is the sum of the weights, and cases with a missing, zero or negative weight are left out — the chip in the header says how many. Every graph says whether the weight is applied to it (a scatter-type plot shows each case once, so its points cannot show the weights). Analyses that compute with the weights take any positive weight; those that repeat each case as many times as its weight (Explore, Kaplan–Meier, Cox, ROC, agreement, clustering…) need whole numbers and say so; the runs test refuses. Every weighted result says “Cases were weighted by w (sum of weights N; k cases excluded)” and its syntax starts WEIGHT BY. Scripts — the Code tab and own-data analysis plans — get the weight variable as a column, dictionary.json names it with any user-missing codes, and the run records the weight; the script decides how to apply it. Frequency weights are not survey weights: for a design-based survey analysis use R's survey package in the Code tab.
  • Split file is an option inside every analysis dialog: the analysis is repeated for each level of the split variable, stacked in one output.
  • Sort cases, Remove duplicate cases (by key), Aggregate (one row per group with n, mean, sum, SD…), Restructure wide → long and long → wide, Merge by key and Append cases. Aggregate, restructure, merge and append create a new derived dataset so the original stays intact.

Transform menu — variables

CommandResult
Compute variableA formula over other variables (ROUND, IF, MEAN, DATEDIFF… — see the function list in the dialog), kept up to date when inputs change
Recode into different variableOld value / range / missing → new code, rules top to bottom, unmatched values copied or dropped
Automatic recodeText categories → integer codes with value labels
Standardizez-scores
Rank cases / N-tilesAverage ranks, or quartiles, quintiles… as an ordinal variable
Visual binningEqual width, quantiles, mean ± SD or custom cut points, with generated labels
Count occurrencesHow many of the chosen variables equal a value, fall in a range or are missing
Indicator variablesOne 0/1 dummy per category
Replace missing valuesMean, median, mode, constant, last observation carried forward or linear interpolation
Lag, cumulative sum, differencesIn case order
String function, Date partUPPER/LOWER/TRIM/LENGTH; year, month, day, weekday, quarter

Analyze menu — the statistics engine

Each analysis opens a dialog: pick the variables by role (only suitable ones are offered), set options, watch the live preview, then “Run and save to Output”. Results come as APA-ready tables with statistics, p-values, effect sizes and confidence intervals, a summary sentence you can paste into a manuscript, and charts where they help.

GroupAnalyses
Descriptive statisticsFrequencies, Descriptives, Explore (CIs, percentiles, outliers, normality, histogram, Q–Q), Normality tests, Crosstabs (χ², Fisher, likelihood ratio, residuals, OR/RR/NNT, κ), Missing value analysis
Compare meansMeans by group, one-sample / independent / paired t-tests (Levene, Welch, Cohen's d, Hedges' g), one-way ANOVA with Welch's F and post hoc tests (Tukey, Games–Howell, Bonferroni, Holm, Scheffé), Bootstrap CIs
General linear modelFactorial ANOVA / ANCOVA with interactions and partial η²; repeated-measures ANOVA with Greenhouse–Geisser and Friedman
CorrelatePearson, Spearman, Kendall matrices with heatmap; partial correlations
RegressionLinear (standardized β, VIF, Durbin–Watson, Breusch–Pagan, residual plots), binary logistic (OR, Hosmer–Lemeshow, classification, AUC), Poisson (IRR, offset, overdispersion), multinomial, ordinal (proportional odds, parallel lines), curve estimation (8 models)
NonparametricMann–Whitney, Kruskal–Wallis with pairwise tests, Wilcoxon, sign, Friedman, McNemar, Cochran's Q, chi-square goodness of fit, binomial, runs, one- and two-sample Kolmogorov–Smirnov, median test, Jonckheere–Terpstra
SurvivalKaplan–Meier (numbers at risk, quartiles, survival at times with differences, the log-rank family, restricted means, late entry, landmark), life tables, Cox regression (strata, start–stop rows, shared frailty, changing hazard ratios, adjusted curves, calibration, case checks), parametric models (Weibull, exponential, log-normal, log-logistic) and competing risks (cumulative incidence, cause-specific and Fine–Gray models)
Scale and agreementCronbach's α with item-total statistics, intraclass correlations, Cohen's κ (weighted), Bland–Altman with Lin's CCC
ClassifyDiagnostic accuracy (Se, Sp, PPV, NPV, LRs, DOR with CIs), ROC (DeLong CI, Youden cut-off), k-means, hierarchical clustering with dendrogram, discriminant analysis
Dimension reductionPCA and principal-axis factoring with KMO, Bartlett, scree plot, varimax rotation, communalities

Assumption checks are built in: t-tests and ANOVA report Levene's test and switch the post hoc test when variances differ; Explore and Normality run Shapiro–Wilk, Anderson–Darling and Lilliefors and say which test family to prefer; regressions report VIF, heteroscedasticity and residual normality.

Regression models — the model builder

Linear, logistic, Poisson, negative binomial, gamma and Cox regression share one way of describing a model, and nothing in it is typed as a formula. Choose the outcome and add the predictors; everything else sits in three sections under them, closed until you want them. A section that is in use says so, so a closed section never hides a choice.

WhereWhat you can ask for
PredictorsA categorical predictor takes a reference level. A numeric one takes a shape: a straight line, a curve (a restricted cubic spline with 3, 4 or 5 knots, or a quadratic or cubic polynomial), the log or the square root of the value.
Model: interactionsLet the effect of one predictor depend on another: pick two of the predictors (three for a three-way interaction; the two-way ones it contains are added for you). A spline can interact with a factor.
Standard errors and confidence levelModel-based errors, errors robust to unequal variance (HC3; HC0 to HC2 are there too), or errors clustered on a variable such as patient or site. The same choice carries through to every test, interval and estimate after the model. The confidence level is 95% unless you choose 90% or 99%.
After the modelFor a factor: its adjusted means (or probabilities, rates) with comparisons between its levels: every pair, each against the reference, a trend, or a contrast of your own, adjusted by Tukey, Šidák, Bonferroni, Holm or the false discovery rate. For a number: its slope, or predictions at values you give, with a plot of its effect. “Within each level of” repeats this inside the levels of a second predictor, which is how an interaction is read. Average effects puts every effect on the outcome's own scale: a difference in mean, in risk or in rate.

The output adds what the model calls for. With an interaction or a curve, a table tests each term as a whole (Type II tests, so a main effect is not judged at the reference level of something it interacts with), a second table asks whether each curve is needed, and the interaction is opened up for you: the means of one factor within the other with their comparisons, or the slope of a number in each group, and a plot with confidence bands. Robust or clustered errors are named in the tables and in the summary sentence.

For this questionChoose
Risk ratios or risk differences instead of odds ratiosBinary logistic regression → Report: risk ratios (modified Poisson with robust errors always fits; log-binomial is the classical model) or risk differences (identity link)
A skewed positive outcome such as cost or length of stayGamma regression: ratios of means on the outcome's own scale, with nothing to back-transform
Counts more spread than Poisson allowsNegative binomial regression, or Poisson with “Quasi-Poisson” ticked, or Poisson with robust errors
Patients measured in several sites, or several rows per patientStandard errors: clustered, with the site or the patient as the cluster variable. With fewer than about 30 clusters the output warns that such errors run small.
Does the treatment work differently in men and women?Add both as predictors, add their interaction, and read “Comparisons of treatment within sex” and the plot
Is the effect of age a straight line?Enter age as a spline and read “Is the curve needed?” and the effect plot

Every number of this part of the engine is checked against R (lm, glm, coxph with the sandwich, car and emmeans packages) and against Python's statsmodels on five datasets that ship with the tests. Robust errors follow sandwich; clustered errors use the same small-sample factors as R and Stata; comparisons follow emmeans.

A saved run writes its whole specification as one syntax line (ANALYZE … interactions=[[arm, sex]] vcov=robust after={…}), and the session's R script writes the same model with lm or glm, sandwich, car::Anova and emmeans, so a colleague can repeat it either way.

Comparing groups and repeated measurements

The General linear model group of the Analyze menu holds the analyses of variance. Each one checks its own assumptions, reports the size of every effect, compares the groups for you, and opens with a reading in plain words.

AnalysisUse it whenWhat you get
One-way ANOVAOne grouping factor, a numeric outcomeF and Welch's F, Levene's test, η² with its interval and ω², post hoc comparisons (Tukey, Games–Howell, Bonferroni, Holm, Scheffé, LSD). Planned contrasts: a trend across the groups, each group against a control (Dunnett), or weights of your own.
Factorial ANOVA / ANCOVASeveral grouping factors, with or without covariates to adjust forUp to five factors with their interactions; Type III tests (as SPSS and SAS) or Type II; partial η² with its interval and ω² for every effect; Levene's test; for an ANCOVA, whether the covariate's slope is the same in every group; adjusted means of every factor with all pairs compared; each interaction opened up; profile plots.
Repeated-measures ANOVAThe same cases measured several times, with or without groupsOne or two within-subject factors laid out over the columns, any grouping factors; Mauchly's test with the Greenhouse–Geisser and Huynh–Feldt corrections; the multivariate test; partial and generalized η²; the means with every pair compared; the change inside each group when the groups change differently; a profile plot.
Mann–Whitney, WilcoxonTwo groups or two measurements, a skewed or ordinal outcomeThe test, the exact p-value for small samples without ties, rank-biserial r, and the Hodges–Lehmann difference with its 95% interval.
Kruskal–Wallis, FriedmanThree or more groups or measurements, a skewed or ordinal outcomeThe overall test with its effect size and the pairs that differ (Dunn's test after Kruskal–Wallis), Holm-adjusted.
  1. Two within-subject factors: choose the measurements in the order where the second factor changes fastest (visit 1 condition 1, visit 1 condition 2, visit 2 condition 1…).
  2. Open “Two within-subject factors”, name the first factor and give the number of levels of the second.
  3. The output starts with a table of how the columns were read. If it does not match your data, reorder the measurements.

When a dialog cannot run yet it says what it still needs (“Still needed: Dependent variable, Fixed factors”). Where a level is typed, such as a reference level or a control group, the dialog suggests the variable's own values.

Factorial ANOVA results saved before version 0.7.1 of the engine tested the main effects of a model with an interaction at a reference level, under the heading Type III. Run them again to get the Type III tests that SPSS, SAS and R report.

Tables of counts, risks and rates

These analyses answer the questions of a clinical or epidemiological study from counts: how much more often an outcome occurs in one group than another, whether that holds across strata, how fast events occur, and whether two treatments are close enough to be called the same.

AnalysisUse it whenWhat you get
CrosstabsTwo categorical variablesThe table with χ², the likelihood-ratio test and Fisher's exact test (for 2 × 2 tables, and for larger tables whenever counts are small). For a 2 × 2 table: the risk in each group, risk ratio, risk difference, number needed to treat, odds ratio (with the exact one), and attributable fractions, read as exposed against reference for the event. With a layer: the Mantel–Haenszel common odds ratio and risk ratio, the Cochran–Mantel–Haenszel test, and the Breslow–Day test that the odds ratio is the same in every layer. With three or more ordered groups: the Cochran–Armitage test for trend.
Binomial testOne proportion against a valueThe exact test and the Wilson, exact (Clopper–Pearson), Agresti–Coull and Wald intervals, with which one to report.
Incidence rates and rate ratiosEvents counted over follow-up timeThe rate per 1,000 (or any number of) person-years in each group with its exact interval, and each group's rate ratio and rate difference against the reference group. Standardise for a second variable (age band, sex) to compare groups of different make-up: directly standardised rates, the Mantel–Haenszel rate ratio, and the SMR (observed events against those expected at the reference rates).
Ordinal regressionAn ordered outcome (mild, moderate, severe)One odds ratio per predictor under proportional odds, with a check of that assumption for each predictor (Brant's test) and the odds ratio at every cut-point of the outcome, so you can see where an effect changes along the scale.
Conditional logistic regressionA matched case-control study, 1:1 or 1:manyOdds ratios that compare each case with the controls of its own set, so the matching factors drop out. Interactions, curved effects, a test of every term and comparisons between levels, as in the other regressions. The output shows how the sets are made up and which carry no comparison.
Equivalence and non-inferiorityThe aim is to show that two groups do not differ by more than a margin, or that one is not worse than the otherTwo one-sided tests against the margin, the interval that answers the question (90% at α = 0.05) beside the usual 95% one, a plot of the interval against the margin, and a verdict in words: equivalent, not equivalent, or inconclusive. For two groups, paired measurements, one group against a value, or two proportions; the margin in the outcome's units, in standard deviations, or as a ratio such as 0.80 to 1.25.
  1. For a 2 × 2 table, check the title of the risk table: it says which group is compared with which, and for which outcome (“yes against no, for event = 1”).
  2. If that is not the direction you want, open “2 × 2 tables: how to read the risks” in the dialog and give the event value or the reference group. The dialog suggests the variable's own values.
  3. Risks, risk ratios and risk differences suit a cohort, a trial or a survey. In a case-control study only the odds ratio has a meaning.

A test that finds no significant difference has not shown that two groups are the same: the study may simply be too small. To claim “no meaningful difference”, decide beforehand how large a difference would matter and run Equivalence and non-inferiority with that margin.

In a matched study do not add the matching variables (the age band, the sex, the clinic) as predictors of the conditional logistic regression. They are identical within every set, so they cannot be estimated; the matching has already held them constant.

Measurements that belong together

When the same people are measured several times, or people sit inside clinics, schools or families, the rows are alike within a person or a site. Treating them as independent makes every test too sure of itself. These three analyses allow for it. They need the data in long format, one row per measurement; Data → Restructure turns a wide file (one column per visit) into a long one.

AnalysisUse it whenWhat you get
Linear mixed modelA numeric outcomeEach person (or site) has its own level and, if you ask, its own rate of change; a second grouping such as the clinic above the patient. Tests of the predictors with Satterthwaite degrees of freedom, how much people differ, whether they do, each person's predicted level, a plot of everyone's measurements under the model's average line, and adjusted means with comparisons.
Mixed model for a yes/no outcome or a countA yes/no outcome or a count, and the question is about one personOdds or rate ratios for a given person, how much people differ (the median odds ratio), a test of whether they do, each person's predicted level.
GEEAny of the three outcome types, and the question is about the average of the whole groupPopulation-averaged effects with robust standard errors, corrected when the groups are few; the working correlation chosen with a table that compares all of them.
  1. Choose the outcome and the predictors as in any regression.
  2. Under “Rows that belong together” choose the variable that marks the person or the site.
  3. If people may change at different rates, open “What differs between the groups” and tick the time variable.
  4. Read “What this means” first: it says how much of the variation lies between people and what each predictor does once that is allowed for.

A mixed logistic model and a GEE answer different questions and give different odds ratios from the same data. The mixed model's are for one person and lie further from 1; the GEE's are for the group as a whole. Report the one that matches your question and say which it is.

Mixed-model results saved before 10 November 2026 used a faulty REML criterion (variances about 1% off), and a GEE's QIC had the wrong penalty. Run them again for the corrected values.

Time until an event

Use these when the outcome is how long something takes and some people have not had it by the end of their follow-up: death, relapse, discharge, a device failing. Each person needs a follow-up time and a variable that says whether the event happened. People who had not had it are censored: they count for as long as they were followed, and no longer.

AnalysisUse it whenWhat you get
Kaplan–MeierYou want the curve, or to compare groupsThe curve with the numbers at risk under it; the time by which a quarter, half and three quarters have had the event; survival at the times you choose with the difference between groups; a test of the curves (log-rank, or one that weighs early or late differences more); the average event-free time; a hazard ratio per group.
Life tableFollow-up is reported in intervals, or all you have is counts per intervalFor each year or month: how many entered, were withdrawn and had the event, the share still event-free at its end, and the hazard.
Cox regressionSeveral predictors at onceA hazard ratio per predictor, each allowed for the others; a check that each ratio holds over the whole of follow-up; survival curves adjusted for the other predictors; how well the model predicts by a chosen time; the cases that pull hardest on a result.
Parametric survival modelYou need a predicted median, or to look past the end of follow-upFour shapes of hazard fitted and compared (Weibull, exponential, log-normal, log-logistic); effects as time ratios; observed against fitted curves.
Competing risksSomething else can happen first and rule the event outThe share who have had the event by each time, the share who had the other outcome first, a test between groups, and each predictor's effect on the rate of the event and on the share who end up with it.
  1. Choose the follow-up time and the variable that says what happened; say which value is the event if it is not 1 or Yes.
  2. Add a grouping variable (Kaplan–Meier) or predictors (Cox, parametric).
  3. Read “What this means” first. Then look at the numbers at risk: a curve is only as firm as the number of people still under it.
  4. Open the optional sections only when your data need them: people who join late, a person with several rows, groups formed during follow-up.
Your dataWhat to do
People join the study at different points of the time scale (age as the time scale; enrolled some time after diagnosis)Give the time each person joined under “When people join late”. They are counted as at risk only from then.
A value changes during follow-up (a complication sets in, a treatment starts)One row per period for each person, each with its start and end, the event on the last row; give the start under “Row starts at” and the person under “Person”. Do not carry the final value back to time 0: someone must survive long enough to get it, and that waiting time would be credited to it.
Groups are formed during follow-up (responded by three months)Set a landmark time. Only people still event-free then are analysed, grouped by what had happened by then.
People come from centres, clinics or familiesFor a few large centres: stratify by centre. For many small groups: “Groups that share an unmeasured factor” (a shared frailty), which also tells you how much the groups differ.
A hazard ratio does not hold over follow-up (the check says so, or the curves cross)Let it change: smoothly with time, or with a value of its own before and after times you choose. Or report the average event-free time from Kaplan–Meier, which needs no such assumption.

A hazard ratio above 1 means the event comes sooner. A time ratio (parametric models) above 1 means it comes later. They point in opposite directions for the same effect; the output says which one it is showing.

Survival results saved before 10 November 2026 should be run again if you report any of these: the interval of a median (its limits were printed in reverse order), the interval of a restricted mean (about 6 to 9% too narrow), numbers at risk at a time point (people whose time was exactly that point were left out), the test of proportional hazards (now the one R prints since 2019), or a hazard ratio allowed to change with time (its test against a constant ratio was not valid when times were tied). Hazard ratios, survival percentages and log-rank tests were right.

Gaps in the data, and treatments that were not randomised

Two problems that a plain comparison hides. Cases with a missing value drop out of a model without a word, and the cases that are left may not resemble the ones that went. And when people were not assigned to a treatment at random, the treated differ from the others before treatment starts, so a difference in outcome is not yet an effect.

AnalysisUse it whenWhat you get
Missing value analysisBefore any model, if some values are missingHow much is missing and in which combinations; a test of whether the gaps are completely at random; for each variable, what goes with its gaps; a plain statement of whether leaving cases out is safe.
A model with multiple imputationLeaving cases out would bias the result or cost too many casesEach gap is filled several times from the other variables, the model is fitted to every filled-in copy and the results are combined. Pooled coefficients and tests, a comparison with the complete cases, a look at what was imputed, and a what-if for gaps that are not at random.
Propensity scoreA treatment or exposure that people were not assigned to at randomA model of who got the treatment; the groups made alike by weighting or matching; balance before and after; the effect with its interval; a second, doubly robust estimate; for a yes/no outcome a risk ratio; for a time to an event weighted curves and a hazard ratio.
Instrumental variablesSomething you did not measure drives both the exposure and the outcome, and you have a variable that moves the exposure and nothing elseThe effect estimated through the instrument, beside the ordinary regression; how strongly the instrument moves the exposure; whether the exposure is confounded at all; whether several instruments agree; an interval that stays valid when the instrument is weak.
Difference in differencesA policy, programme or system reached one group at a known time and not another, and you have the outcome before and afterEach group's change and the difference between them, with errors clustered on the units followed; with several periods the gap between the groups period by period and a test that they moved in parallel before the change.
Regression discontinuityWho gets something is decided by a score and a threshold: a programme from a risk score of 50, a drug below a blood valueThe jump of the outcome at the threshold, corrected for curvature and with a robust interval; the same with a narrower and a wider window; a picture; a test that cases do not pile up on one side; other variables that should show no jump.
MediationYou know that an exposure affects an outcome and want to know how: through an intermediate variable, or by other routesThe effect through the mediator, the effect by other routes, the total and the share mediated, with bootstrap intervals; the two models behind them. Outcome and mediator can each be a number or yes/no.
  1. Run Missing value analysis on the variables of your model. If it says the gaps do not matter, go on as usual.
  2. If they do matter, open “Missing values” in the model's dialog and choose to keep the cases. Add helper variables that go with the gaps.
  3. For a treatment that was not randomised, open Propensity score. Choose the treatment, every covariate measured before treatment that could drive both treatment and outcome, and the outcome.
  4. Choose whom the answer is about, then read the balance table before the effect: an effect from groups that are still unbalanced means little.
The effect in whomThe question it answersChoose it when
EveryoneWhat if all were treated against noneThe treatment could sensibly be given to anyone in the data.
The treatedWhat the treatment did for those who got itYou are judging a treatment as it was actually used.
Those who could have gone either wayThe effect where there was a real choiceThe groups differ a lot. These weights balance the covariates exactly and never become extreme.
The treated, by matchingThe same as the treated, in pairs or sets you can listYou want a matched sample to show. Treated cases with no counterpart are left out, and the output says how many.

A propensity score makes the groups alike only on the covariates you gave it. It does nothing about what was not measured. Say so when you report it, and give the largest standardised difference after adjustment.

A propensity score deals with confounders you measured. An instrument deals with those you did not, but only if it really touches the outcome through the exposure alone. No test can show that for a single instrument: you have to argue it from how the data arose. A random encouragement or a lottery is a strong argument; a convenient variable that happens to correlate with the exposure is not.

Look at the first-stage F before the estimate. Under about 10 the instrument is weak: the estimate drifts towards the ordinary regression and its usual interval is too narrow. Report the Anderson–Rubin interval then, even if it is very wide.

A difference in differences is the effect of the change only if the two groups would otherwise have moved in parallel. Collect several periods before the change if you can: the period-by-period table then shows whether they did. Give the unit that is followed over time (the clinic, the region), or the interval comes out too narrow.

A regression discontinuity is only as good as the rule behind it. If people or their doctors could nudge the score over the line, cases pile up on one side and the two sides are no longer alike: the density check looks for exactly that. The result holds for cases near the threshold and says nothing about those far from it.

Mediation asks more of the data than any other analysis here. Even when the exposure was randomised, the mediator was not: anything that drives both the mediator and the outcome is taken for an effect through the mediator. Add such variables as other predictors, and say in your report that the result rests on having measured them.

If the weighted estimate and the doubly robust estimate disagree, one of the models is wrong. Enter the strongest covariates as curves, add an interaction, and look at the balance again.

Imputed models run before 10 November 2026 should be run again if a variable with two categories had gaps: those gaps were filled at random, whatever the other variables said. The column headed FMI also held a smaller quantity than its name says.

Code — Python and R

The Code tab runs your own script in an isolated sandbox (no network, three-minute limit) against the active cases. Python has pandas, numpy, scipy, statsmodels, lifelines, scikit-learn, seaborn/matplotlib and pingouin; R has survival, ggplot2, lme4, metafor, meta, pROC, car, emmeans, psych, irr, nnet and ordinal.

Python: numbers with hikma.result, tables with hikma.table, figures with hikma.figure(name, fig, caption)

import pandas as pd
d = pd.read_csv(DATA_PATH + "/data.csv")
hikma.result("n", len(d), label="Cases")
hikma.table("means", d.groupby("arm")["sbp"].mean().reset_index(), "Mean SBP by arm")

R: hikma$data() reads the cases; hikma$result / hikma$table / hikma$figure report

d <- hikma$data()
m <- coxph(survival::Surv(time, status) ~ group, data = d)
hikma$table("hr", as.data.frame(summary(m)$conf.int), "Hazard ratios")
hikma$figure("km", function() plot(survival::survfit(survival::Surv(time, status) ~ group, data = d)), "Kaplan–Meier")

Templates in the tab give you a starting script for descriptives, tests, mixed models, Cox models and plots. Every run is saved to Output with its script, so a reviewer can rerun it.

Analysing with AI

Ask your connected agent (project → Your AI) about the dataset in plain language — “does systolic blood pressure differ between arms after adjusting for age?”. Over MCP it describes the dataset and its variable dictionary, runs analyses from the catalogue, writes and runs sandboxed Python or R for what the catalogue cannot express, and reads the results back; every output lands in the Output viewer with the syntax that reproduces it, so its interpretation can be checked number by number.

The Assistant tab's own question box — type a question, the platform plans 1–6 steps, runs them and writes the interpretation from the returned results only — needs platform AI switched on by an administrator. In the default mode the tab explains where to ask instead.

  1. Describe what you want to know, naming the project and the dataset.
  2. The steps are planned against the variable dictionary, preferring analyses from the catalogue over hand-written code.
  3. Each step runs on your real data in the sandbox; nothing is estimated.
  4. The interpretation must come from the returned tables — every number copied, with caveats — because that is all the results contain.
  5. The answer, its steps and their tables are saved to Output and listed under Earlier questions.

From an MCP client (Claude, Codex, Cursor…), the same power is available through describe_dataset, list_analysis_catalog, run_dataset_analysis, transform_dataset, dataset_cases, run_dataset_code and list_dataset_outputs.

From Output into the manuscript

Every saved output has “Insert into manuscript”: choose the manuscript, what to include (the summary sentence, the tables, the charts) and where; the tables arrive as real tables, the charts as figures with captions, all as tracked changes attributed to whoever inserted them — “Claude Code” or whatever your connected agent is called — so co-authors can accept or reject them. Meta-analysis results, risk-of-bias and Summary-of-Findings tables insert the same way from their own pages.

Identifier scan and de-identification

Every dataset is scanned when it is created or imported: variable names, labels and a sample of values are checked for direct identifiers (names, e-mail addresses, telephone numbers, postal addresses, national or hospital numbers, dates of birth, device identifiers). A red or amber banner in the studio lists the flagged columns with the reason. Remove or recode them (Variables tab, Transform › Recode), re-scan, or record an attestation that the data are de-identified — the attestation is kept with the dataset's history with your name and the time. Study codes issued by the project are flagged as “possible” only so you can confirm them.

Graphs — from the dataset, saved as outputs

The Graphs tab (and Graphs › New graph) builds a figure from the dataset's own variables: bar (counts or a summary by category, stacked), line, pie, mean ± SD/SE/95% CI, scatter with trend line, box plot, histogram, Kaplan–Meier with log-rank, bubble, forest and funnel plots. Map the variables and the graph draws at once; style it (title, axes, palette, journal size) and press “Save graph” — it is stored in Output as a chart result with its settings, so you can reopen it with “Edit graph”, export it as SVG or 300-DPI PNG, and copy it to project files. The project's Data › Graphs gallery lists every saved graph across datasets; “New graph” always starts from a dataset. Numbers that are not in any dataset can still be graphed from the gallery's “typed values” panel.

Analyses save their own charts too (histograms with normal curve, Q–Q, error bars, ROC, survival curves, forest plots), and the Analysis hub's “Run an analysis” opens the chosen procedure on the chosen dataset in one step.

Output viewer

  • Pin what matters; rename outputs; delete drafts.
  • Copy any table as tab-separated text for Word or Excel; export any chart as SVG or 300-DPI PNG.
  • Analyses carry a Hikma syntax line (ANALYZE …) — the exact specification that produced them.
  • The datasets page → Outputs lists results across all datasets of the project.

Related

Something missing or wrong? Tell us.