AMOS vs SmartPLS

Two methods behind the two programs

AMOS uses covariance-based SEM, and this method begins from your theory. From your model it builds the covariance matrix the theory expects. Then it compares that matrix with the one in your actual data and tells how close the two are. That is why AMOS gives fit indices, and good fit means the theory is consistent with the data.

SmartPLS uses PLS-SEM, and it works more like many regressions joined one after another. For each construct it makes a weighted score from the items. Then it estimates the paths in a way that explains the outcome constructs as much as possible. It does not try to reproduce the full covariance matrix, so there is no full set of fit indices. Instead you get is R², effect sizes and out-of-sample prediction.

Ask one simple question. Maybe you are testing whether a known model holds in a new setting, for example TAM among farmers in Krishnagiri. Then AMOS is the closer match. If your work is more about which drivers explain an outcome, with a new or extended model, SmartPLS is closer. Hair and colleagues (2019) explain these criteria, and their paper is the usual citation for choosing PLS-SEM.

Your data and your sample size

By default AMOS uses maximum likelihood, and maximum likelihood assumes multivariate normality. So check normality first, before you believe the AMOS output. Go to View → Analysis Properties → Output and tick “Tests for normality and outliers”. The output show Mardia’s multivariate kurtosis along with a critical ratio. A critical ratio above about 5 is commonly read as non-normal data. Likert data from Indian surveys is many times skewed. Even if your data is not normal you need not leave AMOS, because bootstrapping is there inside AMOS also. You will find the Bootstrap tab in Analysis Properties, and it also has the Bollen–Stine bootstrap for the model’s chi-square.

AMOS needs a reasonable sample size also. For a model of moderate size, 200 is the number quoted most, and bigger models need more than that. With 150 responses and 40 items, AMOS estimates may be unstable or may not converge.

SmartPLS does not assume normality. For significance it is doing bootstrapping, so your data gets resampled some thousands of times by the software. A smaller sample can work in PLS-SEM. Still, the idea that PLS works with any sample is wrong. With the inverse square root method, a smallest path of 0.2 needs about 155 respondents. You can go through the calculation in our guide on sample size for a thesis.

What an AMOS analysis involves

In AMOS you draw the model by hand. Constructs are ovals, items are rectangles, arrows are loadings and paths, and every item and outcome gets an error term. Next open Analysis Properties → Output and tick Standardized estimates, Squared multiple correlations and Modification indices. Then run Analyze → Calculate Estimates.

The first model you run in AMOS is the measurement model. This is a confirmatory factor analysis (CFA), where all the constructs are correlated with each other and there are no paths. After the CFA runs, take the fit indices from the output and compare each one with its cut-off. Hu and Bentler (1999) suggested CFI and TLI close to .95 or above, RMSEA close to .06 or below, and SRMR close to .08 or below. Many thesis examiners are accepting CFI and TLI of .90 or more and RMSEA below .08, with CMIN/df below 3. Write the cut-offs you use with their source.

Composite reliability, AVE and HTMT will not come in the AMOS output by itself. You have to calculate them from the standardised loadings. Mostly scholars do it in Excel or with a third-party plug-in. After that you run the structural model with your hypothesised paths. Report the standardised path coefficients along with their p-values.

With modification indices many scholars go wrong, so go slowly there. AMOS will keep suggesting that you free some paths or correlate error terms to get better fit. If two items in the same construct are worded almost the same, joining their errors is sometimes ok and examiners mostly accept it. But if you correlate errors across constructs just to take CFI above .90, that is data-fitting. An examiner who knows SEM will ask you to justify each one of them in viva.

What a SmartPLS analysis involves

In SmartPLS 4 you first import the data. Then you draw the constructs, drag the items on to them and join the paths. Run Calculate → PLS-SEM algorithm to get the estimates. For significance tests, run Calculate → Bootstrapping. Take at least 5,000 bootstrap subsamples. For the final run some newer books are saying 10,000, and on a normal laptop it only takes few minutes more.

Hair and colleagues (2019) give the checks for the measurement model in SmartPLS.

• Each outer loading should be .708 or more, so the item shares about half of its variance with its construct.
• For internal consistency, look at Cronbach’s alpha, rho_A and composite reliability, ideally between .70 and .95.
• AVE of .50 or above shows convergent validity.
• For discriminant validity they use HTMT. Keep it under .85 if your constructs are clearly different in meaning. If two constructs are close in meaning, under .90 is accepted.

The structural model starts with collinearity. Look at the inner VIF values, which should ideally be below 3. Then report path coefficients with bootstrapped p-values or confidence intervals, R² for every outcome and f² for every path. Predictive power you show through PLSpredict. Hair and colleagues call R² values of .75, .50 and .25 substantial, moderate and weak. But what is a good R² differs a lot from field to field.

You will see there is no full set of fit indices in this list. SmartPLS does give SRMR. Researchers are still debating how far SRMR applies to PLS, so if you report it at all, report it carefully.

Fornell–Larcker and HTMT

Discriminant validity checks whether your constructs are actually different from one another. If “trust” and “satisfaction” correlate at .92, your respondents are probably not seeing them as two separate things.

The older test is from Fornell and Larcker (1981). The square root of each construct’s AVE must be higher than the correlation of that construct with every other construct. In the table you put √AVE on the diagonal. It must be the largest number in its row and column.

Henseler, Ringle and Sarstedt (2015) found that the Fornell–Larcker test many times misses real problems. They proposed HTMT, the heterotrait–monotrait ratio, as a better test. SmartPLS gives HTMT directly in its output. In AMOS you can calculate it from the item correlations, or use a plug-in for it. Many theses now report both tests, with HTMT as the main one.

If HTMT comes too high, go back and read the items of those two constructs first. Constructs with almost the same wording will always overlap. Mostly it is better to remove an item that cross-loads than to merge the two constructs. Write in the thesis what you did and why.

Choosing, and defending the choice

A quick way to decide is like this. Go with AMOS if your model is from established theory and you have to show model fit. Your sample should be 200 or more, and normality should be acceptable, or else you plan to bootstrap. Go with SmartPLS if prediction is your main aim, or the model is new, or it has many constructs. Formative constructs, or a smaller and skewed sample, also point towards SmartPLS.

Some departments simply like one tool more. If your guide has worked with AMOS for twenty years, an AMOS analysis may pass your committee more easily. We are not saying ignore that. Only see that the method still suits what your objectives are asking. Discuss about it with your guide before data collection.

Some scholars are coming to us after a bad result and asking to move to the other tool. Say your AMOS model fit is poor and you shift to SmartPLS because SmartPLS gives no fit indices. An examiner who reads your method chapter with the objectives will see why you moved, and it will come up in viva. Poor model fit is also a finding. Explain it in the method chapter, check the measurement model once again, and report the fit as it is.

In your method chapter, give the reason for your choice in two or three sentences with a citation. For PLS-SEM, scholars usually cite Hair and colleagues (2019). Sometimes SEM is more than your objectives need and a regression is enough. Check that in our statistical test guide.

Reporting SEM results in the thesis

Both methods follow the same order. The measurement model comes first and the structural model comes after it. Under the measurement model, give a table of items with their loadings. Then give a table for each construct with alpha, CR and AVE, and after that the discriminant validity table. The alpha column is explained in our guide to Cronbach’s alpha. For the structural model, show a figure with standardised coefficients written on the arrows, and add a hypothesis table. In that table each path gets one row. Put the β there, then the t-value (in AMOS it is the critical ratio) and the p-value, and in the last column write supported or not supported.

In APA style, one path can be written as “Perceived usefulness had a positive effect on intention to use (β = .41, t = 6.12, p < .001), supporting H2.” AMOS gives the critical ratio (C.R.), and you use it in place of t. For AMOS, put the fit indices in one line below the model figure. For SmartPLS, add R² for each outcome and f² for each path.

If you want someone to run your SEM analysis along with you, or check one you already ran, our data analysis service works with both AMOS and SmartPLS.