Reviewers almost always ask why the sample is the size it is. Before data collection the tool computes the sample size required for an assumed effect; after collection it reports the smallest effect the achieved sample could have detected. Nothing is uploaded — it is a calculator.

Power is computed directly from the noncentral t, F and chi-square distributions, covering t tests, one-way ANOVA, ANCOVA, repeated measures, multiple and hierarchical regression, correlation and chi-square. Three tests specific to structural equation modelling are included and are not available in general-purpose power software: the RMSEA-based test of close fit (MacCallum et al., 1996), the inverse square root method for PLS-SEM (Kock & Hadaya, 2018) and joint-significance power for a simple mediated effect (MacKinnon et al., 2002).

The tool deliberately does not compute observed post hoc power. Power calculated after the fact from the observed effect size is a restatement of the p value (Hoenig & Heisey, 2001); when a reviewer asks about power, a sensitivity analysis is the appropriate answer, and the tool routes the user there.

Nearly every empirical paper opens with the same table: who the respondents were, what the items looked like, whether the scales held together, and how the constructs correlate. It is also where numbers are most often mistyped while being copied out of statistical output. The tool builds all four tables from a single raw survey file, and it is method-agnostic — studies that never touch structural equation modelling can use it as it is.

Cronbach's α is computed from its definition. ω (composite reliability) and AVE require loadings, so they are derived from a single-factor principal axis solution; these can differ slightly from a CFA or PLS measurement model, and the report says so and defers to the measurement model when one is reported. The square root of each AVE is placed on the diagonal of the correlation matrix so discriminant validity can be read off the same table.

Items that were never reverse-coded are a common and costly oversight, so they are flagged explicitly: a negative corrected item–total correlation, or an α that rises when the item is dropped, is reported rather than silently absorbed.

Scale development and validation papers report both analyses, which in practice means moving between two statistical packages and copying numbers by hand. The tool computes both from a single raw survey file. For the exploratory part it reports KMO (overall and per item), Bartlett's test of sphericity, and three separate grounds for deciding how many factors to retain: the eigenvalue-greater-than-one rule, parallel analysis, and cumulative variance explained. Extraction is by principal axis factoring or maximum likelihood, with varimax or promax rotation.

Parallel analysis (Horn, 1965) is the default, because comparing observed eigenvalues with the 95th percentile of eigenvalues from random data of the same size does not overestimate the number of factors the way the eigenvalue rule does. When the criteria disagree, the tool says so and leaves the decision to the researcher rather than resolving it silently.

The confirmatory part estimates the measurement model by maximum likelihood, with factor variances fixed at one so that factor correlations are read directly. It reports the usual fit indices with a 90% confidence interval for RMSEA, standardized loadings with significance tests, α, composite reliability and AVE, Fornell-Larcker and HTMT discriminant validity, and the standardized residual correlations that locate misfit. Items that were never reverse-coded appear as negative loadings — their magnitude is normal, so fit indices alone will not reveal them.

A study that measures every construct with one self-report questionnaire has to say what it did about common method bias, and reviewers regularly return manuscripts that leave the question unanswered. This tool takes the raw survey file and runs the three diagnostics that journals in the social sciences ask for.

Harman's single-factor test reports the variance explained by the first unrotated component. Full collinearity VIF regresses each construct on all the others, following Kock (2015). The marker-variable adjustment of Lindell and Whitney (2001) partials the smallest theoretically unrelated correlation out of every construct correlation and re-tests significance, so a correlation that survives only because of shared method variance becomes visible.

The output is a Method-section draft in Korean and English, with the tables, the in-text citations and the reference list already in place. The overall conclusion and the account of procedural remedies are highlighted: neither can be settled from the data alone, and both belong to the author.

Seven designs that all answer the same question — whether groups differ — are gathered on one screen: one-sample, independent-samples and paired-samples t tests, one-way and two-way analysis of variance, analysis of covariance, and repeated measures. Selecting a design reveals only the fields that design needs.

What a reviewer usually objects to is not the p value but what should have been beside it. Effect sizes are therefore always reported: Cohen's d with a confidence interval obtained by inverting the noncentral t distribution, and η², partial η², ω² and Cohen's f for the analysis of variance. Sums of squares are Type III, so unequal group sizes are handled correctly.

What to report when an assumption fails is computed in advance. If Levene's test rejects equality of variances, the Welch test and Games-Howell comparisons are already in the table; if normality is doubtful, the corresponding distribution-free test is too. For repeated measures, Mauchly's test is accompanied by Greenhouse-Geisser and Huynh-Feldt corrected degrees of freedom. The tool also flags reporting an omnibus-nonsignificant post hoc comparison, and interpreting main effects in the presence of an interaction.

Give it a SmartPLS report export, or the raw survey file together with a short description of the model, and it returns APA-formatted tables for the measurement model, discriminant validity, structural paths, indirect effects, explanatory power and out-of-sample prediction, each with the criterion values a reviewer will look for.

Estimation is performed in the tool itself: PLS-SEM and covariance-based SEM by maximum likelihood, bootstrap confidence intervals, specific indirect effects with bias-corrected intervals, and PLSpredict cross-validation. Values that violate a common threshold, and sentences that require the researcher's own judgment, are flagged rather than quietly resolved.

The tables and prose it produces are drafts: figures should be checked against the analysis software, and flagged sentences rewritten by the author.

Mediation and moderation are among the most frequently reported analyses in the social sciences, yet the standard tool for them is a macro that runs on top of a commercial statistics package. The same computations run here: simple and parallel mediation, serial mediation, moderation, and moderated mediation with the moderator on either the first or the second path.

Confidence intervals for indirect effects are obtained by bootstrapping (5,000 resamples by default), reported as both percentile and bias-corrected intervals. Because the sampling distribution of a product of coefficients is not normal, the Sobel test is reported only for reference. Standard errors may be conventional or heteroscedasticity-consistent (HC3), and variables entering a product term are mean-centred.

For moderation the tool reports simple slopes at the mean and ±1 SD of the moderator, solves the Johnson-Neyman boundaries analytically, and states what proportion of the observed range of the moderator the effect is significant over. It also flags two practices that a reviewer will catch: probing an interaction that is not significant, and claiming moderated mediation when the index does not exclude zero.

Where structural equation modelling asks about the net effect of each variable, fsQCA asks which conditions, taken together, produce the outcome, and allows several distinct combinations to do so. Recent work in information systems and marketing increasingly reports both the symmetric and the asymmetric analysis from the same survey file.

The tool calibrates each construct into set membership with the direct method, tests every condition and its negation for necessity, builds the truth table, and minimizes it with the Quine-McCluskey algorithm. Complex, parsimonious and intermediate solutions are all reported, and the configuration table marks core and peripheral conditions in the notation of Fiss (2011).

Calibration thresholds and directional expectations are theoretical judgments rather than properties of the data, so the report leaves those passages highlighted for the author. Configurations whose consistency falls within .02 of the chosen threshold are flagged, because a small change in the threshold would change the solution.

The author supplies two things — the manuscript and the target, either a word count or the name of a journal whose limit the tool knows — and chooses what the journal counts (main text only, with the abstract, or including references). The tool then works in three layers and never rewrites a sentence.

First, wording with a single shorter equivalent is tightened as Word tracked changes under the editor name HJITM: in order to becomes to, empty openers such as It should be noted that are deleted, nominalisations such as conducted an analysis of become analyzed, and quantity phrases, redundant pairs and intensifiers are reduced. Second, sections and tables that need not sit in the main text — common-method-bias tests, robustness checks, non-response analyses, measurement-invariance steps, item tables, cross-loadings, literature-summary tables — are identified from their headings and captions and, if the author ticks them, marked as deleted in the text, replaced by a highlighted pointer sentence, and copied into a separate Supplementary file with formatting, figures and footnotes intact. In-text references to moved tables are renumbered as tracked changes. Third, whatever excess remains is not touched: the report lists where it could come from — hypothesis restatements in the Results, parenthetical statistics already in a table, citation stacks, overlapping paragraphs between the Introduction, Discussion and Conclusion, overlong paragraphs — with an estimated saving for each.

Journal limits are included only where they were read on the journal's own guideline page, with the date checked and a link; page-based limits are converted at an assumed words-per-page rate and flagged for verification. The manuscript is processed in the browser and is never uploaded.

Grammar and awkward phrasing in an English manuscript are corrected and returned as Word tracked changes under the editor name HJITM, so every item can be accepted or rejected individually in the Review tab. Rejecting all of them restores the original file exactly; formatting, emphasis and tables are preserved character by character.

The tool does not rewrite sentences. It changes only what has a single correct answer — subject-verb agreement, count and mass nouns, statistical notation, redundant phrasing, punctuation, and the constructions that recur when a manuscript is translated from Korean. Anything that requires judgment is left untouched and collected in a separate list for the author.

Reference lists are skipped so that bibliographic style is not altered, and paragraphs that already contain tracked changes are left alone rather than being nested inside a second revision layer. The manuscript is processed in the browser and never uploaded, which matters for work that is not yet under review.

Given a manuscript and a target journal, the tool does three things: it applies the journal's formatting, it reconciles the in-text citations against the reference list, and it looks for references that may not exist. Two files come back — the formatted manuscript in English and an inspection report in Korean.

Ten citation styles are built in (APA 7, Vancouver/AMA, IEEE, Chicago, Harvard, Elsevier, Springer, MDPI, Wiley and Taylor & Francis), about sixty journals are matched by name, and anything else is inferred from publisher cues. Pasting the journal's Guide for Authors adds the rules that can be read as numbers: word limits, abstract length, keyword counts, blind review.

Reference verification is the part that matters most. Formatting alone catches malformed DOIs, the example-only prefix 10.1000/, two different papers sharing one DOI, reversed page ranges and impossible years — but it cannot tell whether a reference exists. For that the tool queries Crossref: whether the DOI resolves, and whether the paper it resolves to is the one the manuscript cites. Only each reference's DOI, title, first author and year leave the browser; the manuscript text does not, and the lookup can be switched off. Absence from Crossref is not proof of absence — books, dissertations and reports are often unregistered — so the report states what disagrees and shows the evidence rather than calling anything fake.