Among the great variety of programs (M+, OpenMx, lavaan, sem, Amos, Onyx, ...), what is your favorite program and why? What makes this program better than the others? What features are you still missing?
I am going to give another strong recommendation for Mplus. The program does many kinds of analysis. If you push it very far, say deep into IRT measurment models, it may not have every feature that a dedicated program for that has, but it comes pretty close. That's a lot to expect from one program. More important is how much support there is and the quality of it. There is an excellent users forum that is monitored by the Muthens (the authors) and there is excellent direct support, too. The web site (www.statmodel.com) is full of papers and other useful resources. Nothing else comes close, although other programs have individual strengths. Mplus does not create the path diagrams for you, for instances, whereas, AMOS, for one, does. Mplus gets better all the time, too. They also offer excellent training classes in Europe as well as the US (check the web site). Bob
I don't have a favourite as such, but learnt to use AMOS. Like the fact that you can model transitions between ordinal variables as probit probability curves (using the Bayesian interface). Also like the Bayesian tool for multiple imputations, but it's annoying that you can't use those multiple imputed data directly for a path analysis etc. Would like a simple maximum likelihood / least squares method to handle binary outcome variables, though.
My favourite program for SEM is currently Stata, because it allows me to run data management, descriptive statistics and structural equation models all in one environment. It offers convenient ways to include loops and conditions (e.g., one time using all the data, one time exclude some outliers).
The sem command in Stata 12 (i.e., the current version) has its limitations, though; for example, when it comes to raw categorial data. Mplus is certainly more powerful. Both programs make it rather easy to run SEM, which is nice; but it also increases the temptation to run models without thinking them through. In this regard I sometimes miss LISREL with all its verbose input and matrix definitions.
Using WinBUGS (or OpenBUGS, JAGS, PyMC etc.) has the big advantage that you can specify exactly the model you want, without learning the notations and format requirements of various packages. It naturally allows for non-linear models, too.
The disadvantage is the lack of the tools that help interpreting the results of other software, for example sensitivity analyses have to be programmed, too.
I have used EQS and became somewhat frustrated by the lack of a new promised version. Switched to AMOS where you have the ability to use ordered categories approach to rating scale data using the Bayesian estimation procedure. I found that reassuring as the parameter estimates were pretty similar between the Bayesian and maximum likelihood procedures. Examples and documentation of the AMOS product is excellent. Would like a multi-level facility included at some stage.
I like AMOS, It has good written support materials and examples, which are important when learning a new analysis or program. I agree that a multi-level opition is required.
I also prefer AMOS, because it allows a graphical to model building and does not require thinking in matrices and Greek letters (Lisrel). I also includes routines for invariance tests. However, I am learning M+ for approaching the integration of multilevel modelling with SEM.
I use MPlus, as it is very easy to use, has great support and comes with so many options and strengths. I am also trying Stata 12 SEM module now, as Stata is my main statistical package, but I still like the simplicity of MPlus. On the down side - you can 't create diagrams in MPlus.
I prefer Mplus, as its interface is relatively straightforward (even though it doesn't utilize a GUI) and it has immense power, considerable options for analysis, and a good amount of support materials available.
If you're tied to using a program that offers a GUI (graphical user interface), AMOS is preferred by many. However, I was never able to find a way to implement case weighting in AMOS, which in my area of research is an unacceptable caveat (though the last time I looked for this option in AMOS was about 5 years ago). SAS has recently integrated SEM GUI capability into their JMP interface - I haven't had a chance to try it yet, but it may be a good option going forward for people who are interested in using a GUI-based SEM program.
I use OpenMx because it is integrated in R, very flexible, and free so that there is no hassle with licenses for students. BTW Onyx is an Amos-like front end for the compleat idiot.
Andreas is right. Onyx generates or reads OpenMx code and makes nice pictures. I let my students use it who want to chicken-out of OpenMx. :-) Have not used it myself yet. One does not need to be conversant with R as far as I know.
BTW In OpenMx you can define your own loss function if you want to.
I use MPlus and love it. Very easy to use enables you to edit syntax and the print out is also easy to comprehend. Not many hassles with coding the variables either (unlike STATA). My students also say that it is easier for them. A new version (7.0) just got released in Sept 2012 and available. A new feature is the Mplus Diagrammer that can be used to draw an input diagram, to view a diagram created from an analysis, and to view a diagram created using an input without an analysis. http://www.statmodel.com/verhistory.shtml
Also, I always get help if I can't figure out my mistakes. You can send syntax and they will help/fix it for you. Very friendly and prompt response.
I use AMOS and EQS. In my opinion AMOS it is more user friendly and more graphically appealing. However EQS presents better options to deal with non-normal distribution data.
Initially it is something tedious but eventually it is the most powerful approximation, i.e. to compare different models for the same data. Fundamentally for the philosophy of R, and of the majority of his library, towards the Modeling.
I would like to agree with others who mentioned lavaan (free/open source). There is an active community and the author is very helpful. There are blogs dedicated to support the users. It is very easy to use, and very powerful. You can bootstrap models and you can use dichotomous and categorical data. I used it in parallel with other packages and I am convinced that gives consistent results.
I am going to give another strong recommendation for Mplus. The program does many kinds of analysis. If you push it very far, say deep into IRT measurment models, it may not have every feature that a dedicated program for that has, but it comes pretty close. That's a lot to expect from one program. More important is how much support there is and the quality of it. There is an excellent users forum that is monitored by the Muthens (the authors) and there is excellent direct support, too. The web site (www.statmodel.com) is full of papers and other useful resources. Nothing else comes close, although other programs have individual strengths. Mplus does not create the path diagrams for you, for instances, whereas, AMOS, for one, does. Mplus gets better all the time, too. They also offer excellent training classes in Europe as well as the US (check the web site). Bob
I understand v7 of Mplus now creates path diagrams. They run training courses in Asia too. Mplus has pretty much become the gold standard of SEM/IRT software (IMHO).
That's right. Mplus is great. However, it is expensive. Onyx is a freely available SEM software with a graphical modeling interface (like AMOS and recently Mplus). Since its latest release (yesterday), Onyx can export graphical models to scripts for Mplus, OpenMx, and lavaan.
With regard to Roger's comment, my understand (not experience because I don't have the upgrade) is that the latest version of Mplus, version 7, includes new graphics, including diagrams of the path models. Bob
I agree that MPlus is very powerful, and you has a great followers community that actively posts the the MPlus newsgroups and listserves. I have to admit though, that I am very partial to the GUI approach used in Amos, as well as its integration with SPSS, which is the stat package I regularly use. My guess is you will get many answers to this question, but a lot of it will be driven by what people "grew up on".
In terms of flexibility, I am not sure which program can match Mplus. Mplus is much more than SEM. I recently HAD to learn AMOS and found that it is indeed convenient with simple models. But, when your analysis requires just a little bit more, AMOS quickly loses its advantage over Mplus. In fact, graphic interface may be a good thing with simple models. But with complex models, it is almost impossible to draw a diagram or difficult what the diagram means. But, that could be just me.
lavaan. rapidly growing and very flexible; admittedly, currently not that feature rich as some of the others, but I think that will change soon. My favorite: the mimic-option - I had the pleasure to meet Yves Rosseel last year at a psychometric conferecne, and he explained the laborous endeavour to discover the slight differences appearing in several other programs. Very impressive.
Before, I preferred LISREL. To me, LISREL had one invaluable advantage: I came into contact with the matrices, instead of being pulled away by high-polish GUIs.
AMOS - It took quite a while to get used to some of its quirks and figure out all of the useful features, but I've recently taught others how to use it, and it's really very straightforward. I would recommend it.
There is an impressive list of preferences which has been referred by one contributor as really a function of what you 'were brought up on'. Certainly the investment to learn any of the packages is substantial and this is not just related to the cost of the software (where a licence is required). However what is also very interesting is the difference in values of the parameter estimation. This will vary according to the estimation method employed of course (e.g. ML, distribution free etc.). Any comments on the substantive differences between the same estimation procedures employed by different packages would be fascinating. I note the response of Alexandrowicz Rainer who referred to Rosseel's work on this. Any details would be appreciated.
I personally prefer Mplus for various reasons, some are: 1) It is highly flexible (being syntax based) to perform a whole array of analyses, including item factor analysis, multi-group, and multi-level CFAs, etc., using a wide variety of estimation methods (WLS, WLSMV, ML, MLR, etc.,) 2) The Muthens are extremely helpful, and have a wonderful support site available. Typically they respond to any querries in a matter of days (sometimes even hours), and the discussion boards are rich with interactions that help resolve any issues. 3) The software does not cost a fortune, and is easily available for purchase....
I would say AMOS for most models. I have always loved the book and examples that first came with the 3.6 version. There are some really good support texts that use AMOS, that I like. AMOS does provide more flexibility than people think and is well regarded as a statistical tool in my field of marketing. The downsides are that if can be cluncky at times.
The real key though is not the package, but your understanding of the basis of SEM and the potential pitfalls of this approach. Nothing for me beats a GUI based package btw. I am too old now to learn syntax for new stats programs.
Mplus, Mplus is a flexible program that offers researchers a several models, estimators, and algorithms. Beside Mplus is friendly and an easy to use with the set of nine commands. Mplus is a versatile program with specific capabilities not offered by other programs, including growth mixture, Monte Carlo procedures, three level multilevel SEM (in Mplus v.7) with diagrammer and others.
I have not exposed to Mplus , but, will try later , so far , all the AMOS's features help me a lot for my data analysis , especially , trying to make my R value better ,and close to what the outcomes which I hope it should be , I am quite comfortable with it . But , will try what have recommended by Dr Robert .
In inverse problem exist a forward methods ,it give result from measure and exist inverse problem it give parameter values for completess the structure of equation of problem.
Definitely Mplus. fantastic user support, easy to understand userguide, the website has lots of practical examples and free papers by the authors of the program, there's a large community of users, so you don't have to feel alone with your questions. Most importantly: mplus gives out meaningful warnings when you have made an obvious mistake (model mis-specified, non identification). This makes mplus especially useful for beginners, as it makes you think about your model. Also, mplus is very flexible. R is great too but to me seems to be less stable, needs a lot of memory and is more difficult to learn (I find) and if you want to run a model that does not fit within the available packages, it demands programming skills. There are people who do great things in R in terms of SEM modelling. At my (fairly beginners) level, however, I'm very happy with mplus.
I had used LISREL for many years, now I shifted to lavaan (offering many features similar to M+). But what is most important: You have all your data and your results in one framework. This highly improves the workflow, from accessing the data to publishing the results (with sweave, knitr, R.rsp...).
Rob, I agree. It's a good idea to compare results across programs. It's not only double-checking programs but also double-checking your own programming. Btw., Onyx can export to M+, lavaan, and OpenMx, which is handy for crosschecking results.
AMOS may be user-friendly for a small number of observed items and latent variables, but yo will be out-of-hand if you have over 50 items and 10+ latent variables. I prefer Lisrel and MPlus at this moment. The results shouldn't be different in much if your models are robust.
I think anything related to R is good regarding generating graphs. But I can handle fine with the graphs in Lisrel, but if you are working only with path models with Lisrel, this is the worst program. I don't think MPlus offers any good graphs for you.
Mplus is really good regarding modelling, but the kind of modelling in their graphs is a language for statisticians rather than graphs on causal models.
The most complex model I have ever done with Lisrel is path models with 22 variables. I haven't redone this with Mplus yet. Norammly statisticians would not recommend a model with so many variables with a N smaller than 1000, but we managed with N=300+.
Dear James, Thanks for your input. I welcome the results of your experience. I have purchased STATA 13.1 recently and starting to use the SEM modeller that they have included in their latest version. I would be interested to hear of researcher's views on this packaged product.
I almost had a chance to use STATA when I was still at Nottingham. It is more popular than other packages like Lisrel or Mplus in the UK and it is also similar to SPSS in terms of ease of use. STATA is a comprehensive package that can do MLM and now SEM as you said, while SPSS is separate from AMOS. If you are a beginner, I think STATA is the best, but in the long run, you may need more specialized software. Many people in the UK use STATA to do MLM, but MLWin should be a better and more powerful software.
No serious quantitative researcher can survive with the knowledge of a single package, so getting to know their advantages and limitations as well as your preference and needs is important. I like to work with syntax like Lisrel and equation as in MLWin, so I like them by instinct. I hate AMOS though many people find it user-friendly. I rarely use SPSS myself now and only use it with its syntax. I leave all basic work with my RAs and do only advanced stuff myself.
Many professors don't do analysis themselves now, but I think you have to be an expert first before leaving your crafts. Professors are busy with grant applications and papers. The training on quantitative research in UK universities is far deficient than US ones, so instead of worrying the software package, you should consider training opportunities at other universities where ESRC is sponsoring. And most important, get into a research project as an apprentice asap, you will be paid off in the long run.
What Robert Brennan said is fair. Mplus is really good at support; its weakness is graphs. I am a visual person, so I like graphs, but I don't like drawing graphs. I think in syntax, so both Lisrel and Mplus are fine for me.
Lavaan, because it is an R package (no licensing and open source. There is good support. Can handle categorical data. Makes measuement invariance testing very easy.
James, thanks for taking some time to add to this discussion board. I could not agree more that senior researchers should not leave the factory floor when it comes to data analysis and model testing and still when they get the chance roll up the sleeves and adopt new approaches with various statistical products. I also think that the UK situation as far as training for statistical techniques if far from ideal and not enough investment is made by departments or the individuals themselves to educate and practice good analytical approaches. I was an EQS person but got frustrated by the lack of a new version. I use AMOS as the manuals are good and the Bayesian ordered categorical options very instructive. Can understand however that the software implementation of the diagrammer can annoy.
I think that is no such thing as the 'best" software package that will match everyone's needs. I agree with that a package needs constant updates and support to keep their users, so I want to see Lisrel and Mplus to compete and I can manage both. I absolutely agree with that we also need to keep working on open source like R, but it may not be suitable for everyone. No many people like programming even though it may look very simple to some of us. Back in the 80's we needed to use syntax for SPSS and most of my classmates (particularly female) hated it. The window version may be user-friendly, but very unfriendly if you want to run a similar analysis with very minor changes.
I think Lisrel and Mplus are designed in very different directions. Lisrel is convenient with its SImplus syntax, but its Lisrel syntax is idiosyncratic and may look like working with cryptography. This is why I think Mplus is superior to Lisrel. Its support is also super. However, its modelling is not easy to understand by most people and you will forget about it once you don't use it for a while. I also think that it is easier to handle Multilevel SEM with Mplus.
However, my major concern is that our current training in higher education is so inadequate that many students cannot understand many journal articles. There are also students claiming that they don't have to understand quantitative analysis because they ARE qualitative researchers. I don't think what approach one is adopting for a study doesn't allow him/her to have the privilege to be ignorant of other approaches. This makes me upset when a phd student s/he doesn't have to understand advanced statistics because she is not going to do a quantitative research for his/her thesis. I think Gerry may come across these students quite often.
@ Andreas B. >> one of RG's protocol is that if you ask a question and people take the time to answer it, you should consider 'upvoting' it... @ Robert Brennan and @James Ko , @Gerry seem to have provided very detailed and good answers. Keep it moving. Cheers!
Some weeks ago I had the opportunity to discuss the same topic with Dr. Joe Hair Jr. and his impressions about SEM are very interesting. He is not using LiSREL anymore. On CB-SEM he's using AMOS and he's astronger supporter of PLS-PM SEM. I used all of them and my preferred one is SmartPLS, and it's free!!! (www.smarpls.de)
In general there is not an absolute "best" choice but ideally every different software can be useful for the different features allowed. In this sense exist various software choices to perform the statistical methods. Every software can have as well differences in handling large datasets. So in my opinion, where is possible is useful to compare the results between different software. An additional idea is to use a language as R (and their packages like Lavaan) to integrate the analysis with other methods. For example can be useful during the analysis to detect outliers (and R has some packages to perform this task) and also use the graphical capabilities of the language. Last but not least R it is free so it is possible to compare at no cost the results obtained and expand the analysis.
I used AMOS in the past and moved to mPlus after I encountered serious obstacles in AMOS. More specifically, the problem with AMOS is that you cannot use Satora-Bentler scaled chi, and there is no estimator for categorical outcomes (except for less popular Bayesian approach). MPlus offers both. Besides, what is unique about mPlus is that they offer a remarkably fast and skilled support if you have problems with your model. You send them your code, and they tell you how to make it work. This is extremely helpful. As for the interface, I was at first repeled from mPlus due to its lack of a graphic /drawing mode but after one day of using mPlus I am 100% positive that scripting the model is much easier and faster than drawing / changing it all the time. So the take home message is that given a choice between AMOS and mPlus I would not hessitate to invest in mPlus. I hope my opinion helps!
Thanks Lukasz for your answer which provides some very helpful user experience of the differences between two programmes that have been referred to quite frequently, namely AMOS and MPLUS. You have identified the main criticism that a number of users have complained of, that is the lack of a robust chi - square when quoting results. The catergorical approach works well in AMOS in my experience although it can have a problem with starting values which means that you cannot even start the procedure other than simplify your model.
Opinions welcome of course, especially with concrete examples of the positives and negatives of the various applications that researchers are working with.
If you have to teach SEM to undergrads or to people not familiar with SEM, then you should use AMOS. Its very graphical and easy to use and to explain specification of complex models. Unfortunatelly, SPSS is not upgrading AMOS as it should. It still does not carry Robust estimation methods for ordinal variables like Mplus or EQS (polychorics and tetrachorics correlations), SB corrections, just to name a few, for example. On the open source world I prefer Lavaan. It is pretty close (and easier to program) than Mplus and the Lavaan team is constantyl improving (e.g. they just added invariance analysis with one single command).
amos is much more user friendly, although I find it a little bit too simplistic. I use R, and more specifically the packages sem and lavaan. Lisrel is one of the first programs, it has been around for decades. It is reliable and I think people respect it, but it used to be a little bit difficult to use. I havent used it for years, but in the old days, you had to pre-process the data and then use them. Amos uses data directly from SPSS, so it is the software of choice if you are a beginner.
I agree with João Maroco, the lavaa team (as well as the sem team) in the R world are very energetic. Try R and you cannot go wrong. Right this time I a, writing these lines, I am doing some SEM with lavaan. Very simple, efficient and I only need three commands in total...
Hmmm, Pablo Escribano is right, in the sense that there are so many details one should know before attempting SEM analysis on real data for real-life research. One needs to know some (though not much) information regarding different estimation methods, how ordinal variables need to be treated, what is the impact of severe skewness in the data etc before embarking on serious SEM analysis. Therefore, by the time you get all this info/knowledge, the "user-friendliness" of the software is not important any more. On the contrary, menu-driven packages like AMOS create the illusion that by clicking on a few buttons, anyone can do an SEM. Well, with all due respect, please allow me to disagree.
I am using LISREL software. However, mostly I use the SIMPLIS (Simple Lisrel) syntax base rather than the more complex LISREL syntax base. It is simpler (as the name implies) but not as "user friendly" as the AMOS software.
My vote goes to R (both packages: lavaan and sem). Flexible and feature packed (warning, You have to use the keyboard:). If you insist on using only the mouse, I think that AMOS is the way to go!
My vote goes for MPlus. Having used AMOS (and liked it) as an introduction, I have begun to find it quite limited. MPlus simply does more and at it's current rate of progress, should hopefully continue to do more in the future. This is seemingly unbeknown to many, but the latest versions of MPlus also have a diagrammer, where you can draw and make copies of your model, as in AMOS.
Is there any manual that can help of using AMOS in SPSS?. Iam looking for something like SPSS survival but for AMOS and can take me step by step to how run the SEM.
Perhaps Byrne's 'Structural Equation Modelling With AMOS' book is the nearest you will get? James Gaskin also has some step by step videos on YouTube which you may find valuable e.g. https://www.youtube.com/watch?v=JkZGWUUjdLg
Lavaan package in R have almost everything I need to test and improve the model fit (for example, MI and robust method estimations), and is easy to use, compared to other R packages.
Multivariate analysis in itself is a toolbox. SEM and the related tools for it should always be selected due to a data driven and context specific need. Vendors and specific titles themselves will always change and can be bought or consolidated. Unless the argument of tool selection is driven by parameter estimation needs or algorithmic peculiarities, tool selection is going to be personal preference. The concern with this approach is that as young researchers change institutions there is not a guarantee that the software licensing available at one instruction will be at another.
I currently use AMOS for smaller covariance-based SEM models and LISREL for larger more complex models. I use SmartPLS for all of my variance based SEM. Like any other tool it is context and data specific. I urge people not to fall in love with a specific SEM package and instead understand that they are all just tools.
Stata does pretty much everything that AMOS can do (including the fancy interface, https://www.youtube.com/watch?v=Xj0gBlqwYHI), and most of what MPlus can do except for the more exotic models like latent growth + latent class models, but for the price you are getting a fully integrated package with data management, graphics, and all of other statistics (including biostat, econometrics and survey statistics). Stata 12 was a basic SEM for continuous variables; Stata 13 came in MUCH strengthened with generalized multilevel SEM (binary and ordinal responses, https://www.youtube.com/watch?v=CjsvhOl4ZAw). I would reasonably expect that Stata Corp. will continue developing their capabilities. And if you are still not able to fit a weird multilevel SEM with Stata 13's "gsem" command, there's always "gllamm" as the fallback option (as it has been for the past 10+ years).
If I did not know Stata so well before coming into the SEM area, I would probably be using lavaan in R as one of their developer is a good friend of mine, so I can get a discount on tech support :).