QuickPLS Compare the same study with the same settings
Start with a concrete example. Our SmartPLS comparison uses the same 344 cases, indicator assignments and paths as the published Corporate Reputation model. It compares each named result at three decimal places, the precision available in the source tables.
A method name alone is not enough to establish a fair comparison. Use the same data, variable order, model, missing-data treatment, scaling and case weights. Record the software versions and the choices that affect the calculation.
Choose the outputs to check before inspecting the differences. A path coefficient describes a relationship between constructs; a loading concerns an indicator and its construct. Agreement in one table does not automatically check the other tables.
Ask where the comparison answer came from
A software comparison obtains an answer from another implementation with matching settings. For example, cSEM is an independent R package used in several of our checks. A mathematical check instead calculates the expected answer directly for a defined problem. Both are useful, but they establish different things.
An engine check tests the part of QuickPLS that calculates the numbers. An installed-app check can also examine how the app displays, saves and exports them. A screenshot helps you see the interface; the paired result values show whether the selected calculations agree.
Each comparison identifies its reference and the software that produced the results. Read those details when deciding whether it answers your question. A later app version does not change the identity of a test already completed.
Understand how much difference the test allows
Computers work with a finite number of digits. Two correctly configured calculations can therefore differ slightly in their last digits. Each benchmark sets a rule for deciding whether that difference is small enough. The allowed difference is called a tolerance.
An absolute tolerance is a fixed allowance, useful even near zero. A relative tolerance adds an allowance proportional to the reference value. Many comparisons combine both. Other outputs, such as case counts or whether a result is available, may need to match exactly.
Read the rule beside the reported difference. A number that looks small can fail a strict rule. A comparison of published rounded values answers a different question: it checks agreement at that displayed precision rather than equality of every underlying digit.
On a small screen, swipe the table sideways to see every column.
| Result type | Comparison rule | Why the rule matters |
|---|---|---|
| Deterministic numerical output | The stated absolute and relative error limits. | Finite-precision arithmetic can produce tiny differences between implementations. |
| Results found by an optimiser | The stated error limits and convergence checks. | Stopping rules and parameter definitions can affect a numerical solution. |
| Counts / missing-value patterns / yes-or-no selections | Exact agreement where specified. | An unavailable result has a different meaning from zero. |
| Bootstrap or permutation results | The same sampled rows or a stated statistical comparison. | The same seed number can produce different random samples in different programs. |
Check that the labels mean the same thing
Two tables can use the same label for different definitions. Programs may divide covariance calculations by different values, choose a different sign for a component, or scale scores differently. Match the definitions before judging the numbers.
For example, reversing a whole component's direction can reverse its weights and scores without changing the information it represents. That is different from one unrelated coefficient changing sign. Keep the original outputs and explain any adjustment used to compare them.
A result checked against an independent formula should be described that way. It should not be relabelled as an unchanged output from another software package. The detailed pages name the reference that supports each comparison.
For a bootstrap comparison, match the sampled rows
A bootstrap repeatedly draws data rows and refits the model. To compare each refit between programs, give both the same rows in the same order. For a group permutation, match the reassigned group labels instead. The same seed number in two programs does not guarantee the same random draws.
Check requested, attempted, usable and failed counts. If a comparison uses different random draws, it needs a stated statistical rule for comparing the resulting distributions. Expecting every estimate to be identical would be the wrong test.
Some interval methods use an inner bootstrap within each outer sample. A check of selected inner fits verifies that subset. A check of a summary formula does not establish that every underlying fit was independently repeated. A simulation across invented datasets also needs its data-generating assumptions beside the findings.
Distinguish an invalid input from a calculation disagreement
A good check includes cases where no valid answer should be returned. For example, a model may lack enough independent information to estimate its parameters, or a formula may require division by zero. Refusing such a calculation can be the expected outcome. An unavailable result should not be replaced with a plausible-looking number.
When two valid calculations disagree, keep the difference and investigate its cause. If a later correction or another mathematical check explains part of the issue, identify that separately. The PCA and GSCA pages show how these different findings can be recorded without treating them as the same comparison.
Choose a comparison you can inspect or repeat
To inspect a comparison, open its CSV of named values and read the supporting report. To repeat it, you also need the input data, model, settings and reference software. The detailed page tells you which files are available.
For a downloadable PLS or NCA exercise with checked answers, start with our practice-answer pack. For the direct SmartPLS example, obtain the data from its official source and follow the preparation instructions on the comparison page. Those are practical starting points for checking the same calculation yourself.
Some older records provide saved results rather than a complete rerun kit. Their file fingerprints help identify the original files but do not replace missing data or code. Read the coverage record for that test before planning a reproduction.
- Identify the tested software and the reference used for each output.
- Match the data, model and preparation before comparing values.
- Read the stated precision or allowed numerical difference.
- Keep unmatched outputs and failed fits distinct from successful comparisons.
- Distinguish many values from one model from tests on several independent datasets.