Overview
SEM is often the right tool when your research model has latent constructs (such as trust, satisfaction or self-efficacy) measured by several questionnaire items, and more than one dependent relationship to test. It is also one of the easiest methods to misapply: a model can converge and still be misspecified, an acceptable fit index can hide a poorly measured construct, and PLS-SEM and covariance-based SEM answer slightly different questions.
Our support starts from your theory and research questions rather than the software. We check that your data can support the model you want to estimate, recommend an approach (CB-SEM or PLS-SEM) with a written rationale, assess the measurement model before interpreting any structural paths, and document every decision so you can explain and defend the analysis yourself.
You receive a reproducible analysis — syntax or project files, annotated output and results tables — plus a plain-language explanation of each step. If you want to learn to run it yourself, sessions can be structured as guided walkthroughs instead of a delivered analysis.
Who it’s for and the problems it solves
This is for you if you are…
- Master's and doctoral researchers testing a conceptual framework built from survey data
- Researchers whose supervisor or reviewer asked for measurement invariance, discriminant validity or model fit evidence
- Authors revising a manuscript after comments about common method bias, model fit or PLS-SEM reporting
- Anyone unsure whether their study needs CB-SEM, PLS-SEM or a simpler regression-based approach
Common problems we help with
- The model will not converge, or produces negative error variances (Heywood cases)
- Fit indices look poor and you are unsure which modifications are defensible
- Constructs fail discriminant validity checks such as HTMT or Fornell–Larcker
- You are not sure how to test mediation, moderation or multi-group differences within SEM
- Your results chapter lists output without explaining what it means for your hypotheses
What’s included — and what isn’t
Included
- Review of your conceptual model, hypotheses and questionnaire-to-construct mapping
- Data screening: missing data patterns, outliers, normality, sample size adequacy for the chosen approach
- Recommendation of CB-SEM or PLS-SEM, with a short methodological justification you can adapt
- Measurement model assessment: loadings, reliability (Cronbach's alpha, composite reliability), convergent validity (AVE) and discriminant validity (HTMT, Fornell–Larcker)
- Structural model estimation: path coefficients, standard errors or bootstrapped confidence intervals, R², effect sizes (f²) and, for PLS-SEM, predictive relevance (Q², PLSpredict)
- Mediation (indirect effects), moderation or multi-group analysis where your hypotheses require them
- Common method bias checks appropriate to your design
- Results tables and a path diagram formatted for your style guide (for example APA 7)
- Annotated syntax or project files so the analysis can be reproduced
Not included
- Changing your hypotheses or dropping items purely to improve fit without theoretical justification
- Fabricating, imputing beyond accepted methods, or otherwise altering data to reach significance
- Writing your discussion chapter or interpreting findings on your behalf in submitted work — we explain results so you can write about them
- Guaranteeing that a reviewer or examiner will accept a particular model
What you receive
- Analysis plan
- A one- to two-page plan confirming the approach, estimator, sample-size reasoning and the order of tests, agreed before any modeling begins.
- Measurement model report
- Tables for loadings, reliability, AVE and discriminant validity, with notes on any items that need attention and the options for handling them.
- Structural model results
- Path estimates, confidence intervals, R², f² and hypothesis-by-hypothesis outcomes, plus a publication-style path diagram.
- Reproducible files
- AMOS, Mplus, R (lavaan/seminr) or SmartPLS project files and syntax with comments explaining each step.
- Interpretation walkthrough
- A recorded or live session in which we explain the output line by line so you can describe and defend it yourself.
How it works
Scoping call or brief
You share your research questions, model and a data description. We confirm SEM is appropriate and flag any design issues early.
Analysis plan
We agree the approach (CB-SEM or PLS-SEM), software, estimator and reporting standard, and list any assumptions to test.
Data screening and measurement model
We clean and screen the data with you, then assess the measurement model. Any item or construct issues are discussed before continuing.
Structural model and extensions
We estimate the hypothesized paths, then mediation, moderation or group comparisons if your design includes them.
Reporting and walkthrough
You receive tables, figures and files, followed by a walkthrough and one round of clarification questions.
Methods, estimators and software
- CB-SEM (AMOS, Mplus, R lavaan)
- Suited to confirming established theory when you need global fit statistics (chi-square, CFI, TLI, RMSEA, SRMR) and a well-measured, reflective model.
- PLS-SEM (SmartPLS, R seminr)
- Suited to prediction-oriented research, complex models with many constructs, formative measurement or smaller samples, reported with reliability, validity and predictive criteria rather than a single fit index.
- Estimators
- Maximum likelihood, robust ML (MLR) for non-normal data, WLSMV for ordinal indicators, and bootstrapping for indirect effects.
- Discriminant validity
- Heterotrait–monotrait ratio (HTMT) with bootstrapped intervals, alongside Fornell–Larcker where your field still expects it.
- Measurement invariance
- Configural, metric and scalar invariance (or MICOM in PLS-SEM) before comparing groups such as countries or genders.
- Missing data
- Full information maximum likelihood or multiple imputation where justified, with the rationale documented.
CB-SEM or PLS-SEM: choosing on purpose
Covariance-based SEM tests how well your whole model reproduces the observed covariance matrix, which is why it reports global fit indices. It is the conventional choice when you are testing a theory with established reflective scales and a reasonably large sample.
PLS-SEM estimates the model to maximize explained variance in the outcome constructs. It is increasingly used for prediction-focused and exploratory work and handles formative constructs more naturally, but it should not be chosen simply because a CB-SEM model fits poorly. We help you choose based on your research aims and justify the choice with current methodological literature.
- Theory confirmation with reflective scales → usually CB-SEM
- Prediction, many constructs, formative indicators → often PLS-SEM
- Small sample with simple model → consider whether regression or path analysis is more honest
Measurement comes before structure
Path coefficients are only meaningful if each construct is measured reliably and is distinct from the others. We therefore run confirmatory factor analysis (or the PLS measurement assessment) first and resolve problems such as low loadings, cross-loadings or overlapping constructs before interpreting any hypotheses.
When an item needs to be removed, we document the reason, its effect on content validity and how to report it, so the decision is transparent to your examiner or reviewer.
Mediation, moderation and group comparisons
SEM can test indirect effects with bootstrapped confidence intervals, interaction effects between latent variables and differences between groups. Each needs its own checks — for example, measurement invariance before multi-group comparison — and we set these out in the analysis plan so the tests match your hypotheses.
Limitations we will be clear about
Good fit does not prove a causal model: other models may fit equally well, and cross-sectional survey data cannot establish temporal order. We help you word conclusions accordingly and suggest robustness checks, such as testing an alternative model or controlling for key covariates, where they strengthen your argument.
What the work can look like
Illustrative example — fictional data
Sample deliverable excerpt: measurement model summary
A doctoral researcher surveyed 412 employees to test whether perceived organizational support affects turnover intention through job satisfaction. The figures below are fictional and shown only to illustrate how results are reported.
Reliability and convergent validity
All retained loadings exceeded 0.70. Composite reliability ranged from 0.86 to 0.92 and AVE from 0.61 to 0.72, above the commonly used 0.50 threshold. One job-satisfaction item (JS4, loading 0.48) was removed; the remaining four items still cover the construct definition in the questionnaire source.
Discriminant validity
HTMT ratios ranged from 0.41 to 0.78, and the upper bounds of the bootstrapped 95% confidence intervals were below 0.85, supporting discriminant validity between the three constructs.
Structural result (hypothesis H3)
The indirect effect of support on turnover intention through job satisfaction was β = −0.21, 95% CI [−0.29, −0.14], so the mediation hypothesis is supported. Because the data are cross-sectional, the write-up describes this as consistent with, not proof of, a causal pathway.
Illustrative example with invented numbers. Real deliverables follow your data, your field's reporting conventions and your supervisor's or journal's requirements.
What you’ll need to provide
- Your research questions, hypotheses and a diagram or description of the conceptual model
- The dataset (anonymized) and codebook or questionnaire showing which items measure which construct
- Any required reporting style, template or examiner/reviewer comments
- Software preference or constraints (for example, your university only licenses SPSS AMOS)
- Deadline and whether you want a delivered analysis or a guided learning format
You don’t need everything ready to ask for a quote — tell us what you have and we’ll say what’s missing.
What affects price and turnaround
Price depends on
- Number of constructs, paths and hypotheses in the model
- Whether mediation, moderation, multi-group or invariance testing is required
- Data condition: missing data, reverse-coded items and merging multiple files take extra time
- Delivered analysis versus guided sessions in which you run the analysis
- Depth of reporting support requested (tables only, or tables plus interpretation walkthrough)
Turnaround depends on
- How quickly measurement issues can be resolved with you — they often require a decision on your side
- Availability of a clean dataset and codebook at the start
- Your deadline; urgent work is possible only when the model and data are well defined
Every project is quoted individually. How we price our work.
Questions people ask before starting
How large does my sample need to be for SEM?
There is no single rule. Required sample size depends on model complexity, the strength of loadings, the estimator and the effects you want to detect. We use power analysis or simulation-informed guidance rather than rules of thumb such as '10 cases per item', and we tell you honestly if your sample is too small for the planned model.
My model fit is poor. Can you fix it?
We can diagnose why. Poor fit often comes from measurement problems, omitted paths the theory supports, or a misspecified error structure. We only recommend modifications that you can justify theoretically, and we explain how to report them. We do not add correlated errors or delete items just to reach a cut-off.
Do I have to use AMOS because my university licenses it?
No, but it is often sensible. We can work in AMOS, Mplus, R (lavaan, seminr) or SmartPLS. If a method you need is not available in your preferred software, we will say so and suggest an alternative.
Will you write my results chapter?
We provide results tables, figures and a detailed explanation of what each result means. You write your chapter using that understanding, and we can review your draft for accuracy of statistical reporting. This keeps the work yours and makes you ready for questions at your defense.
Can you check an analysis someone else ran?
Yes. An audit of existing syntax and output is often faster than starting again and is a good way to respond to examiner or reviewer comments.