Adding the moderator construct

Step 1 — Build the WP construct from its indicators. Select WP1–WP3 in the sidebar the same way any other construct is built (click, shift-click, drag to canvas). AnalyVa confirms with “Created WP” and places it as a standalone construct, not yet connected to anything.

Drawing the interaction paths

A moderator needs one interaction path per relationship it’s hypothesized to moderate. This model tests four: does WP change the strength of WM → WSE, WA → WSE, PFT → WSE, or CFT → WM?

Step 2 — H9: WP × WM → WSE. With the Path tool active, click WP, then click WSE. AnalyVa draws the interaction as a dashed purple line — visually distinct from the solid blue structural paths — and labels it automatically.

Step 3 — H10: WP × WA → WSE. Same click-click pattern for the second interaction term.

Step 4 — H11: WP × PFT → WSE. A third interaction term, same pattern.

Step 5 — H12: WP × CFT → WM. The fourth interaction targets WM instead of WSE — the moderator doesn’t have to point at the same endogenous construct every time.

Step 6 — Reposition WP for a readable diagram. With all four interactions drawn, drag WP to a clearer spot (top center, in this case) using the canvas’s arrow controls. All four dashed paths — H9 through H12 — now fan out clearly from WP.

Running the moderated model

Step 7 — Open Run → PLS-SEM → Standard Algorithms → PLS-SEM algorithm again. The same menu used to run the base model in the earlier post — running it again re-estimates with the four new interaction terms included.

Step 8 — Check the moderation-specific setup option, then start. The configuration dialog has the same Path/Factor/PCA weighting choice as before, plus one option that only matters for moderation models: Use SmartPLS-compatible centroid proxy for moderation products. Leaving it unchecked matches R/seminr-style path weighting; checking it matches SmartPLS’s centroid approach for building interaction terms. Pick whichever matches the software you’re benchmarking against, then click Start calculation.

Adding the moderator changes the outcome construct’s explanatory power: WSE’s R² moves from 0.335 in the unmoderated model to 0.366 here — the four interaction terms and WP’s direct effect together account for that gain.

Bootstrapping for significance

Path coefficients alone don’t tell you which moderation effects are real. That requires bootstrapping.

Step 9 — Open Run → PLS-SEM → Bootstrapping.

Step 10 — Read the Beta (p) display. AnalyVa runs 5,000 bootstrap resamples (“fast parallel”) and, by default, annotates each path with its coefficient and p-value inline: -0.034 (p=0.313), -0.053 (p=0.140), -0.007 (p=0.814) — all three visible interaction terms here are non-significant at conventional thresholds.

Step 11 — Switch the Path display to Significance Stars. The dropdown at the bottom toolbar controls how path values render: Beta, Beta (p), Significance Stars, p-value, or None. Switching to Significance Stars trades exact p-values for a faster visual scan — useful once you already know roughly where the interesting cases are.

Step 12 — Read the starred diagram. Every path now carries ***, **, *, or ns, with a legend at the bottom of the canvas (* p<0.05 ** p<0.01 *** p<0.001 ns p>0.05). All four interaction terms read ns. The direct paths — CFT → WSE, WM → WSE, PFT → WSE — stay solidly significant at ***.

Reading the moderation table

Stars on the canvas are a summary. The Results panel has the full picture.

Step 13 — Open the Moderations (Interaction Effects) table. Below the standard bootstrap output, AnalyVa groups the four interaction terms into their own table with a plain-language effect label:

None of the four reach a large effect size, and — cross-referencing the bootstrap table below — none reach significance either (p = 0.610, 0.814, 0.140, and 0.313 respectively). That’s a real, reportable result: this moderator doesn’t meaningfully change any of the four paths tested, at least not in this sample.

Step 14 — Cross-check with the full bootstrap table. Every path — structural and interaction — gets its own row with Beta, SE, t, p, and the 95% bootstrap confidence interval. Hovering any p cell surfaces a plain-language tooltip (“★★★ Highly significant (p < .001)”), which is a fast way to explain a results table to a co-author who doesn’t read t-statistics fluently.

Step 15 — Note which paths are genuinely significant. The base structural paths (PFT → WSE, CFT → WSE, WM → WSE, WP → WSE, CFT → WM, PFT → WM, WP → WM) all clear p<.001. Only the interaction terms are the non-significant group here — a pattern worth stating explicitly in a manuscript: “the direct effects held; none of the four hypothesized moderations were supported.”

Loadings and indirect effects, same run

Step 16 — Scroll to Bootstrap Loadings. The same 5,000-sample bootstrap also re-estimates every outer loading with its own SE, t, and p — every indicator here clears p<.001.

Step 17 — Check Bootstrap Indirect Effects for mediation. Because CFT and PFT both route through WM and WA on their way to WSE, AnalyVa also reports the indirect (mediated) effect for each path — CFT → WM → WSE, CFT → WA → WSE, PFT → WM → WSE, PFT → WA → WSE — with its own confidence interval. This comes free in the same run; no separate mediation analysis needed.

What to report

A null moderation result is still a result. The write-up here would state: R² for WSE improved marginally with the moderator added (0.335 → 0.366), but all four interaction terms were non-significant under 5,000-sample bootstrapping (p > 0.10 in every case), so the hypothesis that WP moderates these relationships is not supported in this sample. That’s a defensible, complete finding — and every number needed to write it came out of a single PLS-SEM run plus one bootstrap pass.

Preparing the grouping variable

The dataset here is a 1,311-row student survey. Gender is already numeric (1 / 2), which AnalyVa can group on directly — but labeling it first makes every downstream table and chart readable.

Step 1 — Open Transform → Recode into Same Variables. With descriptives already reviewed, open the Transform menu. Recode into Same Variables overwrites the values in place — the right choice for turning 1/2 into readable labels without keeping a duplicate column.

Step 2 — Select Gender from the variable list. The dialog lists every variable in the dataset. Click Gender.

Step 3 — Clear the example template. The rules box starts with placeholder text illustrating the syntax (a 5-point reverse-code example, in this case) — not the rule for this variable. Clear it.

Step 4 — Type the actual mapping. One rule per line, oldValue = newValue:

1=Males
2=Females

Step 5 — Click OK. AnalyVa applies the recode in place.

Verifying the recode

Never trust a recode without checking it — a typo in the mapping silently corrupts every downstream group comparison.

Step 6 — Open Analyze → Descriptive Statistics → Frequencies.

Step 7 — Select Gender and run. Tick Include bar chart and Include statistics to get both a distribution table and a visual.

Step 8 — Click OK.

The output confirms the recode worked: 55.2% one label, 44.8% the other, labeled — not 1 and 2 — in every subsequent table.

Running multi-group analysis

With a labeled grouping variable confirmed, the PLS-SEM dialog can split the estimation by group. AnalyVa’s built-in TAM demo dataset already ships with a two-construct model (PU, PEOU) and its own Gender variable, which makes it a fast way to see the grouped-run option end to end — the same Group data sets control works identically on your own recoded variable.

Step 9 — Open Run → PLS-SEM on a model with a grouping variable available. The full PLS-SEM submenu includes both Bootstrap multigroup analysis (MGA) and Permutation multigroup analysis (MGA) as dedicated significance tests for group differences — useful once the basic grouped run below establishes there’s a difference worth testing.

Step 10 — Open the PLS-SEM algorithm dialog. By default, Group data sets is unchecked and the dropdown reads None — the algorithm runs on the full sample.

Step 11 — Tick Group data sets and pick the grouping variable. Once checked, the dropdown lists every eligible categorical variable in the dataset. Select Gender (2 groups).

Step 12 — Confirm both levels and Start calculation. Checkboxes for Female and Male appear, both ticked by default — uncheck either to run a single-group subset instead of a full comparison. Leave both checked and click Start calculation.

Step 13 — Read the result. AnalyVa reports “PLS-SEM done (n=197)” — the complete-case count after listwise deletion — and opens straight into the Smart Model Health panel, which flags model fit and reliability issues before you even look at group-specific coefficients. A group selector (MGA group: All groups) at the bottom toolbar lets you flip the canvas between the pooled model and each group’s own path diagram.

What comes next

A grouped PLS-SEM run gives you two sets of path coefficients side by side, but eyeballing a difference between 0.32 and 0.41 isn’t a significance test. That’s what Bootstrap MGA and Permutation MGA — both visible in the Run menu above — are for: they resample within each group and report whether the path difference survives inference. Run the basic grouped estimation first to see whether a difference is worth testing, then reach for one of the two MGA procedures to confirm it statistically.

Importing the data

Step 1 — Click Import. On a fresh workspace, the canvas is empty and the Results panel just says “Run analysis first.” Click Import in the top toolbar.

Step 2 — Drop the file. The Import Tabular Data dialog accepts .xlsx, .csv, or .tsv.

Step 3 — Check the preview and import. AnalyVa parses the file and reports the shape before you commit — here, 484 rows × 58 columns, all numeric, no missing values. The first six rows are shown for a sanity check. Click Import.

Building a four-construct model

The dataset holds four independent constructs — CFT, PFT, WM, and WA — all predicting a single outcome, WSE. Building each construct is the same click-shift-click-drag pattern used throughout AnalyVa: click an indicator’s first item in the sidebar, shift-click the last, and drag the highlighted range onto the canvas. Repeat for each block, then switch to the Path tool and draw a connection between every predictor and WSE.

Step 4 — The full model with all eight hypothesis paths. CFT and PFT each predict WM and WA directly, and all four predict WSE — eight paths in total (H1–H8). Clicking the WSE construct switches the right panel to Props, listing all 9 of its indicators along with Delete Selected, Select All, and Clear buttons — the same controls used later for indicator cleanup.

Running the algorithm

Step 5 — Open Run → PLS-SEM → Standard Algorithms → PLS-SEM algorithm.

Step 6 — Configure and click Start calculation. The configuration dialog defaults to the Hair et al. recommendations: Path weighting scheme, Standardized results, Mean replacement for missing values (moot here — AnalyVa confirms zero missing cells across the 484 rows). Leave the defaults and click Start calculation.

Reading the Smart Model Health panel

Step 7 — Check the diagnostics before touching the loadings. The Results tab opens with a Smart Model Health summary before any other output. For this run it reads “Review recommended” — no critical failures, but the checklist flags something specific:

Review — Outer loadings. 3 indicator loading(s) between 0.50 and 0.708: CFT / CFT13 = 0.677; WM / WM5 = 0.693; WM / WM12 = 0.699. Suggested action: review AVE and content validity before dropping indicators.

That one line does the scanning work for you — no need to hunt through a loadings table for the offenders. AnalyVa also flags that WA’s R² (0.118) is weak and that bootstrapping hasn’t been run yet, both worth remembering for later.

Step 8 — Find the same indicators on the canvas. With Color-code enabled at the bottom toolbar, loadings below the 0.708 rule-of-thumb render in amber instead of green — the same three values the health panel already named. WM12 is selected here (its loading, 0.699, sits right on the canvas next to it) and CFT13 is outlined in red.

Dropping the weak indicators

Step 9 — Select the offending indicator in the sidebar. Click CFT13 in the left sidebar (or on the canvas). It highlights in red to confirm the selection.

Select the construct on the canvas, switch to its Props panel, tick the indicator(s) to remove — CFT13 and WM12 — and click Delete Selected. AnalyVa updates the construct’s indicator list immediately; no separate confirmation step.

Step 10 — Re-run and check reliability. With the weak indicators gone, re-run PLS-SEM algorithm the same way as Step 6. AnalyVa reports the run as “PLS-SEM done (n=484)”. Hovering a construct now surfaces its reliability block directly on the canvas:

• Cronbach’s alpha: 0.920
• Composite reliability (rho_a): 0.930
• Composite reliability (rho_c): 0.934
• Average variance extracted (AVE): 0.641

All comfortably above the 0.70 / 0.50 thresholds.

Step 11 — Confirm every remaining loading is clean. CFT now runs CFT1 through CFT16 minus CFT13; WM runs WM1 through WM11 minus WM12. Every visible loading is green.

Step 12 — The final model. R² values barely move (WM: 0.211, WA: 0.115, WSE: 0.335 — practically identical to the 12-indicator version), which is expected: two weak formative-adjacent items rarely carry much of a reflective construct’s explanatory power. What changes is defensibility — every loading in the measurement model now clears 0.708 without an asterisk in the write-up.

Why this matters for reporting

Dropping indicators purely to inflate loadings is a real methodological risk — reviewers know the difference between principled cleanup and p-hacking a measurement model. The rule that keeps this defensible: drop an indicator only when its loading is weak and the construct’s AVE improves and the theoretical content coverage doesn’t collapse. AnalyVa’s Smart Model Health panel gives you the first signal automatically; the AVE and content-validity call is still yours to make.

The order that matters

1. Build and run the full model first — don’t pre-emptively drop indicators before seeing real loadings.
2. Read the Smart Model Health panel before scanning tables by eye.
3. Cross-check the flagged indicators against theoretical coverage, not just the number.
4. Delete, re-run, and re-check reliability and AVE — never assume removing one weak item won’t shift another.
5. Run bootstrapping before finalizing path significance (the health panel will remind you if you forget).

This post is the beginner’s tour of that workspace: from opening it for the first time to a fully coded corpus you can query.

Opening the Textual Analysis workspace

Step 1 — Pick Textual Analysis from the home screen. When AnalyVa opens, you get two workspace cards: SEM & Statistics (the familiar quantitative side) and Textual Analysis (the qualitative side). Click Launch Workspace on the second card.

Step 2 — The Textual Analysis workspace loads. The layout is three columns: a Navigation sidebar (Data, Coding, Cases, Notes, Sets, Queries, Visualizations, Analyses), a Files list in the middle, and a large document viewer on the right.

Loading a corpus

Step 3 — Import your documents. Click Import in the top toolbar and select .txt, .pdf, .docx, or .xlsx files. Or, for a first walkthrough, click Demo — AnalyVa ships with a set of 10 policy research documents you can practice on without needing your own corpus first.

Step 4 — Files appear in the Files list. Each imported document becomes a row in the Files list with a created and modified timestamp. In the demo corpus you get ten documents on policy themes: Education Policy, Healthcare Access, Economic Growth, Climate Change, and so on.

Step 5 — Click a document to open it. The right pane opens the document with word and character counts in the top-right, plus four action buttons: Edit, Code Selection, Annotate, Memo.

Step 6 — Skim the text before you start coding. Read the document once end-to-end before annotating anything. Codes built from a first pass through unfamiliar material tend to fragment the theme; codes built from a second pass tend to organise it.

Coding text

Step 7 — Highlight a passage you want to code. Click and drag to select the text — a word, a phrase, a full paragraph.

Step 8 — Click Code Selection in the toolbar above the document. Or right-click the highlighted text and pick Code Selection from the context menu.

Step 9 — The Code Selection dialog opens. It shows the selected text, plus a list of any existing codes you can apply. On a fresh project this list is empty.

Step 10 — Click New Code to create your first code. A prompt asks for the code name — keep it short and descriptive (resource_scarcity, teacher_training, digital_divide).

Step 11 — The code is applied. The passage is now marked with a colored highlight and a small chip naming the code. AnalyVa auto-assigns a colour so each code stays visually distinct.

Step 12 — Apply the same code again to a new passage. Highlight another chunk of text, open Code Selection, tick the box next to your existing code, and click Apply Codes. The passage picks up the same colour as before.

Step 13 — Create additional codes as new themes emerge. Grounded-theory-style coding grows the codebook as you read. New concept? New code. Same concept as before? Reuse the existing code.

Step 14 — Work across multiple documents. Codes are project-wide — a code you created on the Education Policy document is available on every other document in the corpus. This is where the value of a systematic codebook pays off: consistent codes across documents let you query them together later.

Step 15 — Overlap and nesting are allowed. A single passage can carry multiple codes. This is often the point of qualitative coding — a paragraph about “teachers in rural schools” might be tagged both teacher_training and rural_inequality.

Reviewing what you have coded

Step 16 — Open Coding in the left Navigation sidebar. The Coding section lists every code in the project, how many segments it appears in, and across how many documents.

Step 17 — Click a code to see every segment tagged with it. The right pane switches to a Coded Segments view — one row per segment, showing the document, the surrounding context, and links back to the original position. This is the review step where you notice inconsistencies (segments that were mis-coded, codes that should be merged, etc.).

Step 18 — Rename, merge, or delete codes. Right-click a code in the Coding sidebar to rename it (updates every segment automatically), merge two codes into one, or delete a code (with a warning about the segments it will affect).

Step 19 — Add memos to codes and to segments. Codes and individual segments both accept memos — free-text reflection about what you were thinking when you tagged something. Reviewers love these; future-you loves them even more.

Step 20 — Add cases if your corpus has speakers or interviewees. The Cases section links documents to sources (interview respondents, workshop groups, focus groups). This is what lets you later ask “did female respondents talk about X more than male respondents?” — a demographic query on the qualitative data.

Running analyses on the coded corpus

Step 21 — Open Analyses in the sidebar. This is where AnalyVa’s quantitative-qualitative bridge shows up. Every code-count, code-cooccurrence, and code-by-case matrix is one click away.

Step 22 — Run a Codes-by-cases matrix. For every code and every case, the matrix shows how many segments that case contributed to that code. Instantly reveals which codes are concentrated in which respondents, and which are broadly distributed.

Step 23 — Run a Code co-occurrence analysis. For every pair of codes, this counts how often they appear in the same segment — the qualitative analogue of correlation. High co-occurrence between two codes often points to a merger or a higher-order theme.

Step 24 — Open Visualizations for the graphical view. Word clouds, code frequency charts, code hierarchies. Useful in manuscript figures and as a first-look sanity check on the code distribution.

Step 25 — Export. The top toolbar has Export TXT, Export Excel, and Export AVTA (AnalyVa’s own portable format for a full project including codes, memos, and cases). Export AVTA when handing the project to a collaborator; export Excel when moving a code frequency table into a paper.

What to remember

Qualitative coding in AnalyVa follows the same four-stage rhythm as in NVivo, MAXQDA, or Atlas.ti:

1. Import your documents (or use the demo corpus first).
2. Code passages — one at a time, growing your codebook as new themes emerge, applying existing codes to reuse them.
3. Review every code’s segments in the Coding sidebar. Rename, merge, delete as needed.
4. Analyse via Codes-by-cases, Co-occurrence, or Visualizations.

The two things AnalyVa gives you that stand-alone qualitative tools do not: your quantitative variables are still available in the other workspace (switch anytime via Home), and you get an integrated Export AVTA format that packages everything for peer review or collaborator handoff.

AnalyVa handles all of these inside a single Data → Variable View panel — one that will feel familiar if you have ever used SPSS.

Getting to the Data view

Step 1 — Click Data in the top toolbar. Once your dataset is imported, the Data button (rightmost cluster, next to Export) opens a full-screen data browser.

Step 2 — Switch to Variable View. The Data browser has two tabs: Data View (rows and cells, like a spreadsheet) and Variable View (one row per variable, listing type, label, values, missing-value codes).

Every subsequent operation in this post happens in Variable View.

Task 1 — Change the type of a single variable

By default, AnalyVa infers types from the imported data (numeric → Scale, text → Nominal, etc.). To change a variable’s type:

Step 3 — Click the Type dropdown for the row you want to change. You will see three options: Scale (Numeric), Nominal (Categorical), and Ordinal.

Step 4 — Pick the appropriate type. As a quick reminder:

• Scale — interval or ratio (age in years, income, exam score).
• Nominal — unordered categories (gender, country, brand).
• Ordinal — ordered categories with unequal spacing (Likert responses 1–5, education level).

Most survey scale items are technically ordinal, but many analyses treat them as scale in practice. Pick the type that matches the analysis you plan to run.

Step 5 — The change is instant. No Save button, no confirmation dialog. AnalyVa applies the type change immediately.

Step 6 — Sanity-check downstream. Every row shows a small type-icon on the left of the Type dropdown — scale, nominal, or ordinal. Scan the column to catch any variables still typed wrong.

Task 2 — Change the type of many variables at once

If you just imported a 60-item questionnaire, changing types one row at a time is tedious. AnalyVa has a bulk-change mode.

Step 7 — Tick the checkbox on each row you want to change. The leftmost column is a selection checkbox. Ticked rows highlight red.

Step 8 — Pick the new type from Change selected to and click Apply. At the top of the Variable View, next to the Select all checkbox, a dropdown appears with — pick type —. Choose Scale, Nominal, or Ordinal. A counter tells you how many rows the change will affect (here: 11 selected). Click Apply and every selected row updates at once.

Task 3 — Reverse-code a scale

Well-designed questionnaires often include negatively-worded items to catch inattentive respondents. Before analysing the scale, those items must be reverse-coded — a 5 on a 5-point Likert flips to 1, a 4 flips to 2, and so on.

Step 9 — Confirm the recoding applied. The header shows a confirmation like “11 variables set to ordinal” so you know the previous step succeeded.

Step 10 — Return to the main workspace. Close the Data view or press Esc. The sidebar now shows every variable in the dataset.

Step 11 — Open Transform → Recode into Same Variables. The Transform menu at the top of the app holds every data-manipulation command. Recode into Same Variables rewrites values in place — useful for reverse-coding. Recode into Different Variables creates a new column with the recoded values — safer when you want to keep the original.

Step 12 — The Recode dialog opens. Select the variables you want to reverse-code (Cmd/Ctrl-click for multiple, or Shift-click for a range).

Step 13 — Tick Reverse code — or write the rules manually. Two ways to reverse-code:

• Fastest — tick the Reverse code checkbox. AnalyVa infers the scale from the selected variables and generates the rules automatically (e.g. 1=5, 2=4, 3=3, 4=2, 5=1).
• Manual — write the rules directly in the Recode rules box, one per line, using the format oldValue = newValue. Special tokens:
  • LO THRU 3 = 1 — recode “lowest through 3” as 1 (range).
  • ELSE = COPY — keep original for anything not matched.
  • MISSING = 99 — assign a code to blank cells.

Step 14 — Click OK. The recode is applied in place. Back in Data View you can spot-check that the values flipped correctly.

Step 15 — Recompute any composite scores. If you already built mean/sum composites from the raw items via Transform → Compute Variable, they are now stale. Delete the old composites and recompute them so the reverse-coded values propagate. This is the single most common data-prep bug — the reverse-code runs but downstream composites still reflect the un-reversed values.

The order that matters

For any dataset arriving fresh from a survey platform, the safest sequence is:

1. Import the data.
2. Fix variable types in Variable View (bulk-change if possible).
3. Reverse-code any negatively-worded items via Recode into Same Variables + the Reverse code checkbox.
4. Compute composite scores with Transform → Compute Variable.
5. Only then run analyses on the composites.

Skip a step and every downstream result inherits the mistake.

Opening the palette

Press ⌘ + K (macOS) or Ctrl + K (Windows/Linux) from anywhere in the app. A search bar drops from the top. Type the first few letters of what you want to do — bootstr, htmt, fit, pca — and the matching commands filter live. Hit Enter to run the top hit, or use the arrow keys and then Enter.

Close the palette by pressing Esc or clicking anywhere outside it.

Three families of commands

Every command in the palette shows its category underneath the name (Workspace, Analysis, Panels, Canvas). Once you know the categories, guessing what to type becomes easier.

1. Workspace — data in, files out

The top of the palette is your data plumbing. Highlights:

• Import data — ⌘ I — the single most common shortcut. Opens the import dialog. Use it every time you switch datasets.
• Load demo data — for practice. AnalyVa ships with real research datasets so you can rehearse before running your own.
• Open workspace file / Save workspace — ⌘ O / ⌘ S — save and reopen a full model + data + settings state.
• New workspace — ⌘ N — start over cleanly.
• Back to home screen — takes you to the SEM & Statistics vs Textual Analysis picker.

2. Analysis — run the algorithm you actually need

The heart of AnalyVa. Every analytical routine is here — including several that require multi-step menu navigation via the top bar:

• Run PLS-SEM algorithm — ⌘ R — the shortcut you will use the most in SEM work.
• Run Consistent PLS-SEM algorithm — the disattenuated variant.
• Run bootstrapping — for path significance, HTMT CI, and effect size CIs.
• Run CB-SEM algorithm — ⌘ Shift R — for covariance-based SEM.
• Run CB-SEM bootstrapping — bootstrap version of CB-SEM.
• Run SEM regression analysis — for regression-style output.
• Run PCA — principal component analysis.

There are also result-viewing commands under the Results category:

• Show model fit — jumps to the fit indices section of the Results panel.
• Show HTMT — jumps to the HTMT table.
• Show normalized importance — jumps to IPMA results.

If you already ran an analysis, use these to navigate its output without scrolling.

3. Canvas & panels — move around the model

The bottom of the palette is spatial: tools that manipulate what you see on the canvas.

• Hand tool / Path tool — switch tool modes without going to the bottom floating toolbar.
• Residual covariance tool / Factor covariance tool — draw those two special path types for CB-SEM identification tricks.
• Delete tool — delete constructs, indicators, or paths.
• Zoom in / Zoom out / Reset zoom — ⌘ + / ⌘ - / ⌘ 0. Same conventions as your browser.
• Fit model to view — the single most useful spatial command. If your model has drifted off-screen or you are lost in a big canvas, this button reframes everything to fit. Add it to muscle memory.

The productivity pattern

The palette is faster than menus for four specific reasons:

1. It is search-first, not location-first. You do not have to remember which menu contains which command.
2. Every command lives there — including ones that appear only inside submenus of submenus.
3. The keyboard shortcuts are printed on the right of each row, so the palette doubles as a discovery mechanism for the shortcuts themselves.
4. The order is stable. Recently-used commands do not shuffle to the top, so you build muscle memory for the exact position of your most-used commands.

The five shortcuts to memorise this week

If you only remember five, remember these — they cover roughly 80% of routine AnalyVa work:

Do the reps for a week and you will stop reaching for the top menu bar almost entirely.

This post is the short, honest version.

What a p-value literally is

Formally: the probability of observing data as extreme as (or more extreme than) the data you actually observed, assuming the null hypothesis is true.

In plainer language: “if there were no effect in reality, how surprised should we be by our results?” A small p-value means “very surprised.” A large p-value means “not really surprised.”

Notice what a p-value is not. It is not:

• the probability that the null hypothesis is true;
• the probability that your finding is a false positive;
• the size of the effect;
• how important the finding is;
• evidence for the alternative hypothesis in any strict sense.

Any interpretation that starts with “there is only a 3% chance that the result is due to chance” is technically wrong. The 3% is a statement about how unusual the data would be if the null were true, not about how likely the null is.

The five misinterpretations that keep appearing

1. “p = .04, so we reject the null and accept the alternative.” You reject the null. That is not the same as accepting the alternative. Rejection just means the data is unlikely under H₀. The alternative might be true; something else might also be true.
2. “p = .06, so we found no effect.” You found no significant effect at the .05 threshold. The effect could be real and moderately sized — you simply lack the sample size to detect it reliably. Report the effect size regardless.
3. “p = .001, so the effect is huge.” p-values shrink with sample size. A tiny effect in a study of 100,000 people can produce p < .001 while being practically meaningless. Always report and interpret an effect size (Cohen’s d, r, η², f²).
4. “Two studies with p = .049 and p = .051 tell opposite stories.” They tell nearly identical stories. The .05 cutoff is a convention, not a phase transition. Treat the values as continuous.
5. “Non-significant p-values prove the null hypothesis.” Absence of evidence is not evidence of absence. To argue that an effect is truly zero, you need an equivalence test, a Bayesian analysis, or a well-justified prior — not simply p > .05.

What p < .05 actually gives you

It gives you a decision rule that, over the long run, controls your false-positive rate at 5% if all the assumptions of your test hold (independence, distributional assumptions, no p-hacking, no undisclosed multiple comparisons).

Notice how many “ifs” are in that sentence. In real published research, those assumptions are rarely all met — which is why the same p-value carries very different weight in a preregistered replication study versus an exploratory analysis with 30 tested hypotheses.

What to report alongside the p-value

The ASA’s 2016 statement on p-values recommends that authors go beyond a bare “p < .05.” A defensible modern report includes:

• The effect size (Cohen’s d, Pearson’s r, η², or standardised regression coefficient), with its own interpretation.
• A confidence interval for the effect — it carries all the information the p-value carries, plus range.
• The exact p-value, not just “p < .05.” Report p = .034 rather than p < .05.
• A note on the number of tests performed, and any correction applied (Bonferroni, Holm, FDR).

Doing this in AnalyVa

Every inferential test in AnalyVa reports the exact p-value alongside the effect size, group descriptives, an assumption check, and — where relevant — a visualisation. Multiple-comparison corrections are one click away in the same output panel, so nothing about a p-value gets reported without its context.

Here is the full workflow, using an independent-samples t-test as the worked example.

Step 1 — Click Import. Launch AnalyVa on an empty canvas. Click the Import button in the top toolbar (second from the left).

Step 2 — The import dialog opens. An overlay appears asking for a file. Accepts .xlsx, .csv, and .tsv.

Step 3 — Drop your file and preview. Drag your dataset onto the drop zone (or click to browse). AnalyVa parses it and shows the shape and the first rows for sanity-checking (here: 577 rows × 63 columns, all numeric, no missing values). Click Import when the preview looks right.

Step 4 — Open Analyze → Compare Means → Independent Samples t-test. Every variable in the dataset now appears in the left sidebar. Open the Analyze menu → Compare Means → Independent Samples t-test.

Step 5 — The t-test dialog opens. A dialog appears with sensible defaults filled in. AnalyVa auto-detects the grouping variable and its levels — here Gender has two groups, Group A = 1 (n = 234), Group B = 2 (n = 343).

Step 6 — Choose the test variable. Open the Test variable dropdown and pick the numeric variable you want to compare across the two groups. In this example: Teaching_Exp.

Step 7 — Click Run. The dialog now shows your chosen test variable, grouping variable, and both group sizes. Click the red Run button.

Step 8 — Read the results panel. The right pane switches to Results and shows every reporting component in one view:

• Group Descriptives — n, mean, SD, SE per group.
• Levene’s Test for Equality of Variances — F, p, and a plain-English “Equal Var?” verdict (here: Yes, p ≥ 0.05).
• t-test Results — both the equal-variance version (t = -0.582, df = 575, p = 0.560) and Welch’s (t = -0.580, df = 492.71, p = 0.562), with a checkmark on the one Levene recommends.
• Effect Size — Cohen’s d = -0.049, labelled Negligible.
• Box Plot Comparison — a side-by-side visualisation.

Notice how the interpretation lives right next to the p-value. You never have to hunt for the effect size, the assumption check, or the alternative test — everything a defensible report needs is in one scroll.

Every table has its own Copy, APA, HTML, and CSV button above it — so a properly formatted APA sentence (“t(575) = -0.58, p = .560, d = -0.05”) is one click away.

Further reading

• Wasserstein, R. L., & Lazar, N. A. (2016). The ASA’s statement on p-values: Context, process, and purpose. The American Statistician, 70(2), 129–133. DOI
• Amrhein, V., Greenland, S., & McShane, B. (2019). Scientists rise up against statistical significance. Nature, 567, 305–307.

This post gives you a concrete decision path — the same one used by methodologists like Henseler, Ringle, and Sarstedt (2015) and extended in later work. It applies whether you’re running the analysis in SmartPLS 4, in R with the SEMinR or cSEM packages, or in AnalyVa.

What HTMT actually measures

The Heterotrait-Monotrait ratio of correlations (HTMT) compares the average correlation of indicators across two constructs to the average correlation of indicators within each construct. When HTMT approaches 1.0, the two constructs are effectively measuring the same thing.

There isn’t one universally accepted cutoff. The three most commonly cited thresholds are:

• HTMT < 0.85 — strict. Recommended when your constructs are conceptually distinct (e.g., trust vs. satisfaction).
• HTMT < 0.90 — liberal. Acceptable when constructs are conceptually close (e.g., cognitive trust vs. affective trust).
• HTMT confidence interval excludes 1.0 — the inferential test, based on bootstrap resampling. Preferred by many editors today.

If your value sits between 0.85 and 0.90, don’t panic yet — the inferential HTMT test is what most journals now expect you to report anyway.

The six things to try, in order

1. Check the bootstrapped HTMT confidence interval

Run 5,000 bootstrap subsamples with the percentile method and report the 97.5% upper bound. If the interval excludes 1.0, most reviewers will accept the constructs as empirically distinct even when the point estimate is above 0.85. This is the test that survives peer review.

2. Look at the offending indicator pair

HTMT is an average — it hides which specific pair is causing the problem. Pull the raw cross-loadings matrix and find the two indicators with the highest cross-construct correlation. That pair is usually where the theoretical overlap is real.

3. Consider whether the constructs are actually distinct

Sometimes HTMT is telling you the truth: the two constructs aren’t distinct in your data. This is the moment to step back from the statistics and ask a theoretical question. Are perceived usefulness and perceived value really separate for your sample, or did your respondents treat them as the same idea? If the constructs collapse conceptually, merge them and reestimate.

4. Drop the weakest cross-loading indicator

If the pair problem is driven by one indicator with a weak outer loading (< 0.70) and a strong cross-loading, drop it. Document the decision in the methods section. Never drop indicators purely to “fix” HTMT — you need a defensible measurement reason.

5. Reconsider the reflective/formative specification

If one of the constructs is actually formative (its indicators cause the latent variable rather than reflect it), HTMT doesn’t apply in the first place. Formative constructs are assessed with VIF and indicator weights, not with HTMT or AVE. Misspecification is a common hidden cause of “failing” discriminant validity.

6. If all else fails, report and defend

If the theory supports two distinct constructs, the bootstrapped HTMT interval excludes 1.0, and dropping indicators would harm construct coverage, report the HTMT value honestly and defend it in the Discussion. Editors generally accept a defended 0.87 more readily than a suspiciously clean 0.84 that clearly came from indicator surgery.

How this looks in SmartPLS 4

SmartPLS 4 reports the HTMT point estimate in the standard PLS algorithm output, and the bootstrapped confidence intervals under the “Discriminant Validity” tab after you run the bootstrap procedure. To identify the offending indicator pair, open the “Cross Loadings” report.

Doing this in AnalyVa

The full workflow, from a fresh canvas to an HTMT verdict on every construct pair in the model.

Step 1 — Click Import. Launch AnalyVa on an empty canvas. Click the Import button in the top toolbar.

Step 2 — The import dialog opens. An overlay appears asking for a file (.xlsx, .csv, or .tsv).

Step 3 — Load the file and preview it. Drop your dataset. AnalyVa parses it and reports the shape (here: 1,311 rows × 49 columns, all numeric, no missing). The first six rows appear so you can sanity-check. Click Import.

Step 4 — Build the first construct. Every variable now appears in the sidebar. To create a construct, click its first indicator in the sidebar (here, C1).

Step 5 — Shift-click the last indicator to select the range. Shift-click C7 (or Ctrl/Cmd-click individual items). The sidebar highlights the whole range in red. Then drag the selection onto the canvas — AnalyVa creates a construct named after the item prefix (here: C) with all 7 indicators attached.

Step 6 — Repeat for the second construct. The canvas now shows two constructs (C and DP — DP was created the same way earlier). To add the third, click the first indicator of the new construct in the sidebar (here: MC1).

Step 7 — Shift-click through the range. Shift-click MC6 to select all six MC items.

Step 8 — All six items selected. The MC block is now fully highlighted in the sidebar.

Step 9 — Drop the MC construct onto the canvas. Drag the highlighted MC items onto the canvas. AnalyVa creates the MC construct with all 6 indicators. Click the new construct — the right pane switches to Props and shows its type (reflective) and indicator list.

Step 10 — Switch to the Path tool. At the bottom of the canvas is a small floating toolbar. The second button from the left is the Path (P) tool — click it (or press P).

Step 11 — Start drawing a path. With the Path tool active, click the source construct. The cursor is now attached to that construct.

Step 12 — Draw H1: DP → MC. Click the target construct to complete the path. AnalyVa auto-labels it H1 and displays the hypothesis in the header bar (“H1: Path: DP -> MC”).

Step 13 — Draw H2: C → MC. Click C, then click MC. Path H2 is drawn.

Step 14 — Draw H3: DP → C. Click DP, then click C. All three hypothesis paths are now drawn. The header confirms “H3: Path: DP -> C”.

Step 15 — Open Run → PLS-SEM → Standard Algorithms → PLS-SEM algorithm. The model is ready to estimate. Open the Run menu → PLS-SEM → Standard Algorithms → PLS-SEM algorithm.

Step 16 — Configure and Start calculation. A configuration dialog opens: weighting scheme (Factor / Path / PCA), initial weights, missing-value handling. The defaults are the Hair et al. recommendations (Path weighting, mean replacement) and AnalyVa also reports the current data profile — 1,311 rows, 28,842 indicator cells, 0 missing. Leave the defaults and click Start calculation.

Step 17 — Read the HTMT table. The right pane switches to Results and streams every report the Hair et al. checklist requires. The HTMT table sits high on the page with the verdict already labelled:

• DP – C: 0.4180 → Good
• DP – MC: 0.4821 → Good
• C – MC: 0.7090 → Good

All three pairs are safely below 0.85. Outer VIF, Fornell-Larcker, cross-loadings, and the path diagram with computed loadings (0.827, 0.791, …) and path coefficients (0.384, 0.555, 0.228) are all generated in the same run — nothing to re-execute.

If any HTMT pair here had exceeded 0.85, the workflow to fix it starts right back at the top of this post: check the bootstrap CI, find the offending indicator pair in the cross-loadings table below, and decide.

Validation

AnalyVa’s HTMT calculations were benchmarked against SmartPLS 4 on the corporate reputation dataset (n = 344) as part of the second validation study. Point estimates and bootstrapped confidence intervals reproduce SmartPLS output to the reported precision. See the study on Zenodo at doi.org/10.5281/zenodo.21536948.

Further reading

• Henseler, J., Ringle, C. M., & Sarstedt, M. (2015). A new criterion for assessing discriminant validity in variance-based structural equation modeling. Journal of the Academy of Marketing Science, 43(1), 115–135. DOI
• Franke, G., & Sarstedt, M. (2019). Heuristics versus statistics in discriminant validity testing. Internet Research, 29(3), 430–447.

Here is the decision framework and the reporting standards.

The three t-tests

One-sample t-test. Compares one group’s mean to a known reference value. Example: are the students in this sample scoring above the national average of 100?

Independent-samples t-test. Compares the means of two different groups. Example: do men and women differ in exam scores?

Paired-samples t-test. Compares two measurements from the same subjects (or matched pairs). Example: did the students’ scores change from pre-test to post-test?

Using the wrong one inflates or deflates the Type I error rate. An independent test on paired data throws away statistical power; a paired test on independent data is simply invalid.

Decision tree

• Comparing one group to a fixed number? → one-sample.
• Comparing two independent groups? → independent.
• Comparing two measurements from the same subjects (or matched pairs)? → paired.
• Comparing three or more groups? → not a t-test. Use ANOVA.

Assumptions

All t-tests assume:

• the dependent variable is continuous (interval or ratio);
• observations are independent within each group;
• the dependent variable is approximately normally distributed within each group.

The independent-samples t-test additionally assumes equal variances between groups (homogeneity). When that is violated, use Welch’s t-test — most modern statistical software defaults to Welch’s for exactly this reason. It costs nothing when variances are equal and works correctly when they are not.

Effect size — report it, always

A p-value tells you whether a difference exists. Cohen’s d tells you how big it is.

• d = 0.20 → small
• d = 0.50 → medium
• d = 0.80 → large

A study with n = 1,000 might return p < .001 for d = 0.05 — statistically significant, practically meaningless. Always report d alongside its 95% confidence interval.

Three common mistakes

1. Running pairwise t-tests on three or more groups. Three groups A, B, C tested with three t-tests (A vs B, A vs C, B vs C), each at α = 0.05, gives a family-wise error rate of roughly 14%. Use ANOVA followed by a post-hoc test (Tukey’s HSD, Bonferroni) instead.

2. Ignoring Welch’s t-test. When group variances differ notably (Levene’s test p < 0.05), the classical Student’s t-test is biased. Welch’s has no such assumption. Modern journals expect Welch’s by default when variances are unequal.

3. One-tailed tests without prior justification. A one-tailed test doubles statistical power on one side, but requires you to have committed to a direction before looking at the data. Post-hoc one-tailed tests are a form of p-hacking. Report two-tailed by default unless you have a preregistered directional hypothesis.

What to report

For an independent-samples t-test, a defensible modern report:

“The intervention group scored significantly higher than the control group, t(48) = 3.21, p = .002, d = 0.91, 95% CI for d [0.34, 1.48].”

Components:

• t(df) = value — the statistic and its degrees of freedom.
• exact p — not just “< .05”.
• Cohen’s d plus its 95% CI.
• The mean difference plus its 95% CI where journal style allows.

Doing this in AnalyVa

The workflow is almost identical for all three t-test variants — the only real decision point is which one you pick from the menu in Step 4. This walkthrough uses an Independent Samples t-test as the worked example, and calls out at each step how the other two variants diverge.

Step 1 — Click Import. Launch AnalyVa on an empty canvas. Click the Import button in the top toolbar.

Step 2 — The import dialog opens. An overlay appears asking for a file (.xlsx, .csv, or .tsv).

Step 3 — Load and preview. Drop your dataset. AnalyVa reports the shape and shows the first rows so you can confirm it parsed correctly. Click Import.

Your dataset needs to include:

• One-sample: the continuous variable of interest;
• Independent-samples: a continuous outcome + a grouping variable with exactly two levels;
• Paired-samples: two continuous variables measured on the same cases (pre + post, spouse-A + spouse-B, etc.).

Step 4 — Pick the right t-test from Analyze → Compare Means. This is the only decision that changes. Open the Analyze menu → Compare Means. The submenu lists all three t-tests explicitly:

• One-Sample t-test — for comparing one group’s mean to a fixed reference value.
• Independent Samples t-test — for two different groups (this example).
• Paired Samples t-test — for two measurements from the same subjects.

If your design matches multiple items, re-read the decision tree earlier in this post — only one of them is appropriate for any given research question.

Step 5 — The test dialog opens with defaults. Pick Independent Samples t-test and AnalyVa opens a dialog with sensible defaults filled in. Here it auto-detected Gender as the grouping variable and mapped its two levels (Group A = 1, n = 234; Group B = 2, n = 343). The two group sample sizes appear as a one-line summary so you can catch obvious data errors immediately.

For the other two variants the dialog looks slightly different:

• The One-Sample dialog asks for a Test value — the reference number to compare the mean to (population norm, chance level, etc.).
• The Paired-Samples dialog asks for Variable 1 and Variable 2 — the two paired measurements on the same cases — instead of a test variable + a grouping variable.

Step 6 — Choose the test variable. Open the Test variable dropdown and pick the numeric outcome you want to compare across the two groups. In this example: Teaching_Exp.

Step 7 — Click Run. The dialog now shows your chosen test variable, grouping variable, and both group sizes. Click the red Run button.

Step 8 — Read the results panel. The right pane switches to Results and shows everything a defensible t-test report needs, in one view:

• Group Descriptives — n, mean, SD, SE for each group.
• Levene’s Test for Equality of Variances — F, p, and a plain-English “Equal Var?” verdict. Here: Yes, p ≥ .05.
• t-test Results — both the classical Student’s version (t = -0.582, df = 575, p = .560) and Welch’s (t = -0.580, df = 492.71, p = .562), with a ✓ on whichever one Levene’s test recommends. You never have to remember which to report — AnalyVa flags it.
• Effect Size — Cohen’s d = -0.049, labelled Negligible. No hunting for the effect size in a separate output.
• Box Plot Comparison — a side-by-side visual so a reviewer can see the distributions at a glance.

For a paired t-test the same output appears with an extra Correlation between paired scores row and the mean of the differences instead of the mean difference between groups. For a one-sample test, Levene disappears (only one group) and the results compare the group mean to the reference value you specified.

Every table has Copy, APA, HTML, and CSV buttons. Click APA on the t-test Results table and it copies a ready-to-paste sentence like:

“t(575) = -0.58, p = .560, d = -0.05”

into your clipboard, ready to drop into the manuscript.

Further reading

• Lakens, D. (2013). Calculating and reporting effect sizes to facilitate cumulative science: A practical primer for t-tests and ANOVAs. Frontiers in Psychology, 4, 863.
• Delacre, M., Lakens, D., & Leys, C. (2017). Why psychologists should by default use Welch’s t-test instead of Student’s t-test. International Review of Social Psychology, 30(1), 92–101.

r is bounded between -1 and +1. But knowing that is not enough to interpret it responsibly.

What r actually measures

r is a scaled version of covariance. It measures the linear relationship between two continuous variables. Its full name — the Pearson product-moment correlation coefficient — is worth remembering because every word matters.

• Linear. r only detects straight-line relationships. A perfectly curved relationship (U-shape, quadratic) can produce r ≈ 0.
• Product-moment. it is built from the joint deviations of both variables from their means.
• Coefficient. it is a one-number summary — not a full picture. Always plot the data.

Interpreting r’s magnitude

Cohen’s benchmarks (1988) are the most widely cited:

• **

  r

  ≥ 0.10** → small

• **

  r

  ≥ 0.30** → medium

• **

  r

  ≥ 0.50** → large

These are field-specific. In physics, r = 0.90 is common. In social psychology, r = 0.30 might be the strongest effect you will ever see. Compare to your field’s typical effect sizes, not to the benchmarks alone.

Four things r does not tell you

1. Causation. Perhaps the most-repeated caveat in statistics. A high r between ice cream sales and drowning deaths does not mean ice cream causes drowning — a third variable (summer weather) drives both.

2. Non-linear relationships. A dataset with a perfect U-shape can have r = 0. Always inspect a scatterplot before trusting r. If the relationship is non-linear, transform the variable (log, square root, quadratic term) or use a non-parametric measure like Spearman’s ρ.

3. The magnitude of change. r = 0.80 means the variables move together strongly, but it does not say by how much y changes when x changes by one unit. For that, you need the slope of a regression (β), not r.

4. Robustness to outliers. A single extreme point can pull r from 0.10 to 0.60. Always compute r with and without any influential outlier, and report both — or use a robust correlation (Spearman, Kendall’s τ) if outliers are theoretically meaningful and should stay in.

What to report

• The value of r to two decimal places, e.g. r = 0.42.
• The exact p-value, e.g. p = .003 — but read the P-values post first.
• The 95 % confidence interval for r (via bootstrap or Fisher’s z-transformation).
• The sample size n.
• A scatterplot showing the actual data.

Doing this in AnalyVa

A realistic workflow: many correlations in applied research are not between two raw variables but between two composite scale scores (e.g. the mean of six items measuring a construct). This tutorial covers the whole path — data import → build two composites via Transform → Compute Variable → run Correlate → Bivariate → read the result and the heatmap.

Step 1 — Click Import. Launch AnalyVa on an empty canvas. Click the Import button in the top toolbar.

Step 2 — The import dialog opens. An overlay appears asking for a file (.xlsx, .csv, or .tsv).

Step 3 — Load and preview. Drop your file. AnalyVa parses it and reports the shape (here: 361 rows × 53 columns, 49 numeric, 0 missing). Click Import.

Step 4 — Open Transform → Compute Variable. Every variable now appears in the sidebar. Before correlating, build the two composite scores. Open the Transform menu → Compute Variable….

Step 5 — Name the first composite. The Compute Variable dialog opens. In Target Variable, type a short, meaningful name — here Variable1 — this becomes a new column in your dataset.

Step 6 — Move to the Expression field. Click into the Expression field. Below it, AnalyVa shows every available variable as a clickable chip and the full list of supported functions and operators.

Step 7 — Enter the composite formula and click OK. Type or click-to-insert the arithmetic that produces the composite — here the mean of six items: (GP_1+GP_2+GP_3+GP_4+GP_5+GP_6)/6. Click OK. The header confirms “Computed ‘Variable1’ for 361 cases”.

Step 8 — Repeat for the second composite. Open Transform → Compute Variable… again to build the second composite. Same dialog, same flow.

Step 9 — Name Variable2. In the new dialog, type Variable2 in Target Variable.

Step 10 — Build the expression by clicking chips. Instead of typing, click the variable chips (PA_1, PA_2, …) below the expression box — AnalyVa inserts them at the cursor. Add the + between each and wrap in parentheses.

Step 11 — Confirm the second composite. The full formula: (PA_1+PA_2+PA_3+PA_4+PA_5+PA_6)/6. Click OK. The header confirms “Computed ‘Variable2’ for 361 cases”.

Step 12 — Open Analyze → Correlate → Bivariate. With both composites in the dataset, open the Analyze menu → Correlate → Bivariate….

Step 13 — Select the two composites and choose Pearson. The Bivariate Correlations dialog opens with every numeric variable listed. Ctrl/Cmd-click Variable1 and Variable2 to select just the two composites. Leave Method on Pearson. Click OK.

Step 14 — Read the results. The right pane switches to Results and shows a Bivariate Correlations (Pearson) table:

• Variable1 ↔ Variable2 = -0.332 ***

Below the table, a Correlation Heatmap visualises the magnitude and direction (red = negative, blue = positive). Significance stars follow the convention * p < .05, ** p < .01, *** p < .001.

In plain reporting: the two composites correlate r = -0.33, p < .001 — a small-to-moderate negative association. To also inspect whether the relationship is linear (Cohen’s caveat #2 above), open the scatterplot from the same panel before writing up.

Further reading

• Rodgers, J. L., & Nicewander, W. A. (1988). Thirteen ways to look at the correlation coefficient. The American Statistician, 42(1), 59–66.
• Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd ed.). Erlbaum.