A result is reproducible when the underlying data, configuration, and documented procedure produce the same number independently.
This page lists the available outputs, what each contains, and what to check when re-deriving a reported figure.
Respondent-level data
| Output | Contents |
|---|---|
| SPSS export | Full .sav with its Excel codebook as a ZIP bundle, in wide or stacked, coded or uncoded form |
| Wide Excel | One row per respondent. Raw wide exports include non-complete respondents |
| Long Excel | Stacked format, one row per response |
| CSV | Raw or stacked, for direct load into any analysis environment |
Coded exports return recoded categories; uncoded exports return the underlying response values. Use those when checking how a reported category was constructed. Where a report nets or recategorizes responses, comparing the uncoded export against the report is how you confirm the mapping.
Because raw wide exports include respondents who did not complete, they also support checking terminations and break-off directly rather than inferring them from the achieved sample.
Verifying the sample itself
Respondent reconciliation downloads, scoped to the selected dataset, reconcile respondents against the respondent source — the basis for checking delivery against invoice, and for confirming which respondents were excluded and at which stage. Read alongside the live incidence breakdown of completed, in progress, terminated, and poor quality per source described in Respondent sourcing methodology, this makes the quality pipeline's exclusions visible rather than merely asserted. The pipeline itself is documented in Data quality methodology.
Respondent transcripts export the survey as the respondent experienced it, the fastest way to check that routing, piping, and question wording behaved as specified.
Verifying the analysis
Weighting. The diagnostics report the efficiency ratio and the distribution of weights, and the export carries the respondent-level weights themselves, so weighted results can be recomputed independently from the raw responses. Which stages ran, on which universe, against which targets, is part of the configuration rather than a black box. See Weighting methodology.
Analytic percentages and means. Match the report's filters, response recodes, numerator, and denominator base. For Response frequency and Uncoded frequency, inference applies to each category's share of the selected percentage base. Reproduce respondent-level Taylor contributions using the final weights, combining repeated records within respondent and retaining zero contributions outside each cell's domain. The variance correction and test degrees of freedom use the full filtered respondent population. Comparisons require joint covariance; the marginal errors alone are insufficient when cells share respondents. For means without outliers, also reproduce the retention rule and treat the retained set as fixed. See Weighting methodology.
Discrete choice. For conjoint and MaxDiff, individual-level utility scores and raw Hierarchical Bayes draws support offline analysis. Match the question, dataset, calibration method, retained draws, reporting weights and cell population to the report being checked. Use draws from the same calibrated or uncalibrated model as the report.
Recompute each transformation within each respondent and draw: normalize attribute ranges for importance, normalize item utilities for simulated share, and count utilities strictly above zero for Probability above anchor. Marginal Purchase Lift additionally needs the purchase-intercept draws and the same empirical profile counts; giving every profile equal weight changes what the result measures. You need those inputs and the individual draws to reproduce the calculation; mean utilities alone are insufficient. See Utility and simulated share methodology for the formulas.
To reproduce discrete-choice errors and comparisons, align the draw identities across cells, calculate the average within-draw respondent Taylor covariance, and add the between-draw posterior covariance. Preserve overlap between items and respondent groups. The joint matrix supplies both marginal errors and contrast variance.
Lift. The report exposes its own specification: which question or tag is the group indicator, which values count as treated, each outcome's calculation and cut-by source, the matching questions, the number of matching iterations, and the p-value threshold. With the respondent-level export and that specification, the exposed and untreated bases and the unmatched outcome rates can be recomputed directly. Match the outcome question identity as well as its display label, apply the outcome-eligibility and brand-specific control rules, and reproduce the selected respondent groups for each matching run. Selection is deterministic given the same respondent data, configuration, and reporting implementation.
For significance, derive one binary outcome per respondent and metric, average their contrast coefficients across valid runs, and reproduce the conditional score test within the matching strata. Respondents absent from a run contribute zero. The public mean_p_value field contains the combined score p-value. Use unrounded calculations when checking the p-value filter. Respondent coefficients and outcomes, diagnostic standard errors, and diagnostic interval bounds are internal calculations excluded from public lift results and response schemas. Independent verification therefore requires reproducing the matching procedure and inference from the underlying data. See Lift measurement methodology.
Significance. Match the uncertainty procedure to the output: Taylor contrasts for analytic percentages and means, paired respondent bootstrap comparisons for medians, joint sampling and posterior covariance for discrete-choice summaries, or conditional score inference for matched Lift. Apply the configured confidence level, minimum cell count, or p-value filter as appropriate. Lift's Mean P-Value displays the combined score p-value; a 0.95 p-value threshold is a broad filter, not 95% confidence. No automatic multiple-comparison adjustment is applied. See Understanding stat testing and Sample size and precision.
Programmatic access
Surveys, structure, and cross-tab reports are available through a public API with OAuth authentication and through an MCP integration for agentic analysis, so verification can be automated or run outside the interface. Feasibility is exposed the same way. See Public API for reporting and agentic analysis and MCP and OpenAPI integration.
What to check first
If you are auditing a single reported figure, the fastest route is usually this order: confirm the base the figure rests on and whether it is weighted or unweighted; pull the uncoded export and confirm the response values behind any netted or recategorized number; check the weights and the efficiency ratio; then reproduce the point estimate, its standard error, and the configured comparison or p-value filter.
Most discrepancies resolve at the first or second step, and they are usually definitional (a different base, or a different treatment of partial completes) rather than computational.
Scope
Exports contain what the study collected. They do not contain the respondent-source-side data the platform never receives, such as a sample provider's own profiling history or recruitment records; where a study needs that, it has to be agreed with the provider directly. Personal data in exports is whatever the study chose to collect: MX8 Labs always stores a hashed IP address and a unique respondent identifier, and anything beyond that is the study's decision. See Privacy compliance and, for current certification status, trust.mx8labs.com.