Latest Version of NGENE

[tn73]

I would like to find out if NGENE 1.1.2 is the latest version and whether or not a newer version is likely to be out soon. I recall that some 2-3 years ago, experimental designs for latent class models were not available and that some work was being done along these lines.

Thanks
Tom

[paulm]

Seeing no official reply, I'll take a stab at it. I haven't seen any activity on the development front for Ngene in years. I would doubt there is anything pending in the short term.

In other words, don't hold your breath.

[Michiel Bliemer]

Version 1.1.2 is the latest official release, and since last year we have sent an unofficial version (to be called version 1.1.3) to a select number of users that have asked for specific functionality regarding reading in candidate sets, which makes the software more versatile (i.e. it allows more complex constraints, partial profile designs, and availability designs).

Optimalisation for latent class models can more or less be done with the current version of Ngene. To this end, you need to specify a separate model for each class (each with their own prior values) and then optimise the design over all classes under the assumption of some properties of each class appearing in the population. For example something like:

;eff = 1*classA(mnl,d) + 3*classB(mnl,d) + 2*classC(mnl,d)

;model(classA):
U(A) = b1[0.2]*x1[1.2.3] + b2[-0.3]*x2[1.2.3] + b3[0.1]*x3[-1,0,1] /
U(B) = b1*x1 + b2*x2 + b3*x3

;model(classB):
U(A) = b1[0.3]*x1[1.2.3] + b2[-0.4]*x2[1.2.3] /
U(B) = b1*x1 + b2*x2

;model(classC):
U(A) = b2[-0.1]*x2[1.2.3] + b3[0.2]*x3[-1,0,1] /
U(B) = b2*x2 + b3*x3

In this example, not all attributes appear in all classes, and the parameter values are assumed different for each class. The weights 1, 3, and 2 indicate the weights within the population, namely class probabilities of 17%, 50%, and 33%. it is not possible to include a class assignment probability based on covariates, which are usually not available anyway at the time of survey design. The above approach is an approximation of optimisation for a latent class model that you may wish you use.

It is true that developments in Ngene are not going fast as we were hoping, we have focused mainly on bug fixes. Version 1.1.2 can do what most people are asking for, while version 1.1.3 (we are happy to share the test version to users that require the above functionality) will add more features.

Releasing new versions is simply a matter of time and cost. We chose for a business model in which users only have to pay once and we have been giving software updates for free for the past 7 years. If we would have chosen for a business model with annual fees, which would generate much more revenue but also would make the software more expensive, then we would have the means to hire more software engineers. We have chosen for the first option to keep Ngene accessible to a larger audience, including academics and PhD students. We also chose to run this forum along side to assist as many people as possible in designing surveys and experimental designs. In that sense we are not running ChoiceMetrics as a commercial entity, but rather our aim is to assist the (mostly scientific) community in conducting these experiments. We hope that people using Ngene understand and appreciate this choice, which also means that we are not able to release software updates like other more commercial software vendors.

Having said that, our aim is still to release version 1.1.3 officially this year with the above mentioned capabilities. Thank you for your patience.

Michiel

[Nassar]

Michiel

"Releasing new versions is simply a matter of time and cost. We chose for a business model in which users only have to pay once and we have been giving software updates for free for the past 7 years. If we would have chosen for a business model with annual fees, which would generate much more revenue but also would make the software more expensive, then we would have the means to hire more software engineers. We have chosen for the first option to keep Ngene accessible to a larger audience, including academics and PhD students."

Adding features to NGene needs times and ressources, some of us (users) would need Availability Design, other Best-Worst, Menu choices..
Just a simple suggestion.. Is it possible to consider having serveral additional modules the interested people can buy?

Best
Nassar

[Michiel Bliemer]

It is an interesting idea, although we typically prefer providing all functionality in one product. The question is on what functionality we should focus first.
We are always looking for feedback from users on what functionality they would like to see implemented first. There are many extensions that are possible to add, and you name just a few. Most demand has been for availability designs, partial profile designs, overlap designs, and a more robust handling of contraints in Ngene. Therefore, we are currently focusing our efforts on these, and version 1.1.3 should enable this.

Version 1.1.3 (currently in beta stage and is not yet supported, but can be obtained upon request via contact@choice-metrics.com):
Enables reading in candidate sets created by the user (in e.g. Excel). Such user defined candidate sets allow the creation of more complex designs, for example (i) availability designs (in which only a subset of alternatives are presented), (ii) partial profile designs (in which only a subset of attributes are selected), and (iii) overlap designs (in which a subset of attributes have the same level). All these enable simplifying the choice task. With this version one can also include more complex constraints, since the user can create a set of feasible choice tasks in Excel that Ngene can read in. Example candidate sets in Excel show how to create such designs. We are also working to specify availability designs and partial profile designs directly in the syntax, although clearly the most flexible way is by letting the user provide the candidate set.

It would be interesting to hear from others which functionality one would like to see next? Best-worst? Menu choice? Other?

[paulm]

I have a request for the next version. Currently, Ngene is TERRIBLE at blocking. There doesn't seem to be much attention paid to blocks and how they affect results, but I believe for random effects models version effects are large and an important thing to minimize. Ngene could be much better at blocking. One way to do it is to create a design first and then block it in a second step. Another may be to increase the time for blocking in a way that actually works. Another might be to develop better criteria to assess the quality of blocks. I bet you could get several papers out of this topic if you invest some time.

Paul

[Michiel Bliemer]

Dear Paul,

You will be pleased to hear that we have indeed spent quite a bit of time on improving blocking in Ngene.
Note that formally blocking does not exist for efficient designs, in the literature it is only discussed for orthogonal designs and you can use blocking with orthogonal designs in Ngene if you worry about blocking (although I am not convinced blocking can have large effects on model estimation). We believe that some sort of 'blocking' is also useful for efficient designs, but there is no theory or literature on this (and no other software that I am aware of is able to block efficient designs). We chose to minimise correlations between the blocking column and the other design columns, which mimics blocking in orthogonal designs and typically leads to a reasonable degree of attribute level balance, but cannot guarantee it. Therefore last year we tested several other measures, including measures that directly consider level balance instead of indirectly via correlations. No measure is perfect though, so if you have suggestions then please email us at contact@choice-metrics.com.
Ngene already does for certain model types what you suggest, namely it first creates a design and then it blocks it in a second stage. Note that it is more appropriate to include blocking in the design optimisation, in particular for random effects models, since a different random draw needs to be done for each block, while the same random draw is required within each block in order to appropriately take panel effects into account.

Michiel