Statistics and interview coding in one project.
This browser’s Pro key was not accepted: the subscription behind it has ended. Renew or subscribe below.
Research onlyUS$2 a month, 7 days free EverythingUS$5 a month, 7 days free Sign in
US$2 a month or US$19 a year for Research, or US$5 a month or US$48 a year for everything. 7 days free. Stripe takes a card, and nothing is charged if you cancel before they end.
Taken from OBSCURA Research itself, with the demo project it opens in one click: an invented evaluation of a community garden programme, with a survey of 60 respondents and two interviews. Nobody in it is real.
SPSS’s tests and NVivo’s coding in one project, with the data and the transcripts kept on this device. Nothing is uploaded. Open .sav, Stata, CSV and Excel files; run t-tests, ANOVA, MANOVA, regression, mediation, factor analysis, survival and the non-parametric tests, with APA tables; code interviews, recordings and open answers, query them, and measure agreement between coders. For students, academics, NHS and charity researchers and market researchers.
Before you start: The common SPSS tests, not all of them: mixed models with one grouping factor, and no structural equation models beyond confirmatory factor analysis. What else it does not do
A data view of cases and variables, edited in place, with value labels shown or not (Ctrl L), and a variable view of each variable’s name, type, width, decimals, label, value labels, missing values (up to three, or a range and one more), display width, alignment and measure. Filtered cases are struck through, as SPSS shows them.
Opens SPSS .sav files (uncompressed, compressed and .zsav), Stata .dta files from Stata 8 onwards, CSV and Excel workbooks. Saves .sav, .dta, CSV and .xlsx, and the whole project, both sides, as one .oresearch file. The .sav and .dta readers and writers are tested both ways against pyreadstat, the Python reader built on ReadStat.
Compute with SPSS’s own expression language (MEAN.4(sat1 TO sat5), RND, MISSING, IF, dates such as DATEDIFF and XDATE.YEAR, LAG and seeded random numbers), recode into the same or a new variable, automatic recode, count values within cases, rank cases, visual binning, select cases by a condition or at random, identify duplicates, split file, sort, weight cases, merge files by cases or by a key, aggregate, restructure from wide to long and back, reverse-score items, standardise, make dummy variables, and replace missing values, singly or by multiple imputation (chained equations, pooled by Rubin’s rules).
Frequencies, descriptives, means by groups, explore (the 5% trimmed mean, percentiles, Shapiro-Wilk and Kolmogorov-Smirnov with Lilliefors’ correction, extreme values and outliers by the boxplot rule) and crosstabs with chi-square, the likelihood ratio, Yates’s correction, Fisher’s exact test for any size of table and Cramér’s V. Missing value analysis: the patterns, Little’s MCAR test and EM estimates.
One-sample, independent (with Levene’s test and Welch’s correction) and paired t-tests, each with its Bayes factor if asked (the JZS prior, as JASP); one-way ANOVA with Welch’s and Brown-Forsythe’s tests, ω², and Tukey HSD, Games-Howell, Bonferroni, Šidák and Scheffé comparisons; factorial ANOVA and ANCOVA with Type III sums of squares, estimated marginal means and Bonferroni, Šidák or Tukey comparisons; repeated measures with Mauchly’s test and the Greenhouse-Geisser and Huynh-Feldt corrections, and mixed between-and-within ANOVA; MANOVA with Pillai’s trace, Wilks’s lambda, Hotelling’s trace, Roy’s root and Box’s M; Pearson, Spearman, Kendall and partial correlation.
Linear regression in blocks or by stepwise selection, with R² change, standardised betas, VIF, residual plots, each case’s Cook’s distance, leverage and studentised residual, and robust standard errors (HC0 to HC3, or clustered); linear mixed models with random intercepts and slopes, by REML or maximum likelihood, with Satterthwaite’s degrees of freedom; generalised linear models: Poisson, negative binomial, binomial with logit, probit or complementary log-log links, and gamma, with offsets; Breusch-Pagan, White and RESET tests; cluster analysis by k-means or by Ward’s or other linkage, with a dendrogram and silhouette widths; binary, multinomial and ordinal logistic regression, with the test of parallel lines; ROC curves with the area, DeLong’s interval and the best cut-off; mediation and moderation as in PROCESS models 4 and 1, with bootstrap confidence intervals, conditional effects and the Johnson-Neyman interval; Kaplan-Meier with the log-rank, Breslow and Tarone-Ware tests, and Cox regression.
Mann-Whitney, Wilcoxon, Kruskal-Wallis and Friedman with Dunn’s pairwise comparisons; chi-square goodness of fit, the binomial test, McNemar and Cochran’s Q. Cronbach’s alpha and McDonald’s omega with item statistics; agreement by Cohen’s kappa (weighted or not), Fleiss’s kappa, Krippendorff’s alpha and the intraclass correlation. Factor analysis by principal components, principal axis or maximum likelihood, rotated by varimax, quartimax, promax or oblimin; confirmatory factor analysis with CFI, TLI, RMSEA and SRMR; power and sample size for the t-tests, ANOVA, regression, chi-square, correlation and two proportions.
Tables in APA style, charts (histogram with the normal curve, boxplot, scatter with its fit line, bar chart with 95% error bars, Q-Q plot, scree plot, profile plot), a note on each test in plain words and a sentence to report it. Everything run is also written as SPSS syntax (T-TEST GROUPS=group(1 2) /VARIABLES=wb_post.), which can be edited and run again, with TEMPORARY honoured; a colleague’s .sps file opens into it. The output goes to Word, PDF, HTML, or Excel with a sheet for each table.
Word documents, the text of PDFs, plain text and Markdown come in as sources; pictures are coded by drawing a region; audio and video are coded by their times, beside a transcript lined up with the recording (WebVTT, SRT or timestamped text, or typed here with keys to play, pause and stamp the time). A survey’s open-ended columns become sources whose every answer belongs to its respondent. Codes form a tree with colours and descriptions, and can be merged, moved and renamed. Memos go on the project, a code, a source or a passage, and annotations on any span of text.
Search the text and code every hit (or its sentence or paragraph); word frequency with English stop words and Porter’s stemmer; keyword in context; coding queries such as "Cost" AND NOT "Access" or "Weather" NEAR/10 "Travel"; matrix coding of codes by an attribute; co-occurrence as a heat map; framework matrices of cases against codes, with a summary written in each cell beside its passages; word clouds; and counts by case sent to the data as variables, or a matrix sent straight to a chi-square test.
Each coder works in their own copy of the project. Merging a colleague’s file brings their coding in under their name, and Reliability gives Cohen’s kappa, Krippendorff’s alpha and percentage agreement for every code, character by character, as NVivo calculates it. Projects go to and from NVivo, ATLAS.ti and MAXQDA as REFI-QDA files, recordings and their coding included.
An analysis or an importer somebody has written, installed from a file. Each runs in a sealed box on the device: it is given only the variables it asks for, or the one file it reads, and the browser lets it reach nothing, no network and no storage. What it hands back is shown as tables and text.
Every statistic, degree of freedom, p-value and effect size is computed in this page, in JavaScript written for it, and compared on every run of the site’s checks with SciPy and statsmodels on several data sets, normal and skewed, with ties and with missing values. They agree to six significant figures; the largest difference in the last run was a few parts in ten million. The distributions are checked too, including the studentised range that Tukey’s and Games-Howell’s tests use. Mixed models, generalised linear models, robust errors and ROC curves are checked on published data sets as well (lme4’s sleep study, the RAND Health Insurance Experiment, Grunfeld’s investment data, Hanley and McNeil’s CT ratings), against the figures their authors printed; iterative fits agree to five significant figures, and a mixed model’s variances, where the likelihood is flat, to four. Where SPSS has its own definition (skewness and kurtosis, percentiles, Levene’s test, Type III sums of squares) that is the one used, and where it differs from SPSS the note under the table says so.
Interview transcripts, recordings and survey answers often hold personal data, and ethics committees and data protection rules usually say where they may be kept. OBSCURA Research is a page that runs in your browser: a project is kept in this browser’s private storage on this device, files you open are read here, and files you save are written here. Nothing is sent to a server, and there is no account, no licence server and no fee for each seat. Once opened it keeps working without a connection.
For an ethics application or a data protection impact assessment: the data, transcripts and recordings are processed only on the researcher’s device, and no third party receives them. The page asks this site whether its key is still paid and counts the page view (what each request carries); none of those requests carries data, a file or a file’s name. Your data protection officer can check this in the browser’s network panel.
Said now, so that nobody pays to find out.
Data, sources, codes and output together, kept in this browser on this device.
An invented community garden evaluation: a survey of 60 with open answers, and two interviews.
SPSS .sav, Stata .dta, CSV, Excel, or a project saved from here.
Code the open answers, each tied to the respondent’s other answers.
Type data in, or bring transcripts in for coding.