jasp-stats / jasp-issues Goto Github PK

Not sure if this is the intended behavior, but the sorting of the Included Variables in the EFA Factor Loadings table is not ideal. It's neither numerical, nor faithful to the original data file. Screenshot is attached where the included variables are named Q1-Q141. The first three variables in the output are Q1, Q10, and Q100, instead of Q1, Q2, Q3.

Environment (for bugs)

• JASP version: 0.8.2.0
• OS name and version:
• Analysis: Bayesian t-tests

Environment (for enhancement requests)

• Enhancement:

• Add statistics in Bayesian t-test tables: Column "t", column "df"

• Purpose:

• June

Hi

I was wondering whether the RMSEA values I receive are incorrect due to an additional zero decimal.

For example, I receive the RMSEA value of 0.007, 90%CI [0.007 - 0.091]. I introduce a new variable (only 1) to the model and the new RMSEA is 0.079, 90% CI [0.071 - 0.083]

I suspect that the extra zero decimal in the first value is a bug.

Thank you for your contributions to the software.

• JASP version: 0.8.1.1
• OS name and version: Win 7
• Analysis: Exploratory Factor Analysis

From Twitter: Right now, most of our analyses show 95% credible intervals (in the prior-posterior plots). Sometimes people might want a different interval, an x% one, where they get to specify x. We could work that in. Perhaps it would need to go under "advanced options" or something, where we offer the user more flexibility concerning the plotting functionality.
E.J.

Currently, the Bayesian regression only displays Bayes Factors. To really understand the model(s), I'd like to see the parameter estimates and their credible intervals. In a way, this is simply an extension of #1718 to Bayesian regression. The BayesFactor package supports this.

A few more ideas that I just want to drop here, but which are lower on the priority list:

• A checkbox for prediction intervals would also be nice for Bayesian and frequentist regressions alike.
• The ability to set priors on each covariate (lmBF supports this via the rscaleEffects argument). Too bad it doesn't support it for slopes, though!

Currently using JASP 8.1.1. Thank you so much for working on this!

In a two-way ANOVA with two factors (A, three levels, and B, two levels), a planned comparison should be available for both variable A (for example A1 vs. A2 + A3), variable B (B1 vs. B2) and variable A within a single level of variable B (A1 vs. A2 + A3 within level B1).

Environment (for bugs)

• JASP version:
• OS name and version:
• Analysis:

Environment (for enhancement requests)

• Enhancement: Simple linear regression Case Diagnostics (or similar) to provide confidence intervals and prediction intervals
  
  SPSS gives this under the Save (!) option calling them Mean and Individual prediction intervals:
  

• Purpose:

Hi!

I wondered if it possible to include the option of specifying hypotheses within the Bayesian approach of ANOVA. I would like to perform a bf analysis for multiple comparisons ('(linear) contrast analysis').

Thanks a lot for this program! I like it really much!

Hi there and congratulations for JASP! It is wonderful! I already use it with my undergraduate students.

How about considering to implement in JASP, robust analyses developed by Rand Wilcox (http://dornsife.usc.edu/labs/rwilcox/software/)? Robust analyses are extremely useful in psychological research, where in many cases the assumptions of classic parametric tests are violated. These analyses are already available in R: https://cran.r-project.org/web/packages/WRS2/index.html

All the very best,
Konstantinos Priftis

Environment (for bugs)

• JASP version: 0.8.1.1
• OS name and version: Mac Sierra
• Analysis: Descriptive
  Load the debug dataset, select first contNarrow, then contWide. Click on Display correlation plot: 'This analysis terminated unexpectedly'
  What strange is, if you select fist contWide, and then contNarrow, you don't get any error...

As of now, JASP knowns nothing about times and dates: These are converted to a string on import.
Given that databases, SAV files, R and Qt all have date and time support, we might at least attempt to format time and date according to the user's locale - even if no analysis expressly requires it.

A nice feature might be to able to code your own contrast. SPSS calls this the LMATRIX in ANOVA designs (or MMATRIX in repeated measures designs). In R there is a function called fit.contrast() in the gmodels package. This would allow estimation of an interaction contrast. Currently, only pre-build contrasts are possible.

Would also be highly interested in a Bayesian counterpart. Using the BayesFactor package it is easy to get credible intervals for interaction contrasts, but unsure how to get a Bayes Factor for one contrast of interest.

Probably simple to implement and helpful for those who work with weighted data: Weight by weighting variable for descriptive output (specifically Descriptive Statistics and Frequencies).

A suggestion through twitter:
"Perhaps u need a "restore defaults" button? Changed my prior parameter, dunno what default value was pic.twitter.com/xckuVRqoQR"

From the twitterverse:
"Sorry, but that's not fully APA-formatted. Column headings only capitalize first letter, & title of table is italicized."
@raviselker Let's discuss whether we want to do this, because it might conflict with our other formatting conventions

In order to make factor loading tables easier to interpret, particularly with large clusters of variables, I think it would be useful to add a "sort by loadings" option. The function would probably use the fa.sort function of the psych package. Thanks!

Environment (for bugs)

• JASP version:
• OS name and version:
• Analysis: Graphing with CI or SE and effect sizes are important but it might also be helpful to include power analysis to support hypothesis tests.

I just noticed that the repository size increased by 100MB since v0.8~beta2. This is primarily due to TestFiles.zip. Please consider using git-annex for such files (http://git-annex.branchable.com). The inflation of the repository with binary file updates would otherwise lead to usability issues over time.

Any sense in having the total N or degrees of freedom in correlation plots? Also, is it worth having a way to compute simple correlations without having to do a matrix?

Hi,

OLS regression is not fit for variables that are not normally distributed, e.g. dummy or censored variables.

Since many variables of interest fall into that category, it would be very useful to have logit, tobit and probit functions in addition to OLS.

Cheers!

http://psycnet.apa.org/journals/xge/141/1/2.pdf

see different effect size measures. There is growing demand for these, in one way and two way anovas, in t tests, in within and between subject designs

Environment (for bugs)

• JASP version:
• OS name and version:
• Analysis:

Environment (for enhancement requests)

• Enhancement: Descriptive statistics for multiple response sets.
• Purpose: Many surveys include multiple response questions that need to be calculated. I think it would be very useful to include some enhancement at least for descriptive statistics for this type of data.

Hello, I'm having a problem when I import a data file with twitter ID numbers they are transformed into a five digit number with e-17 trailing. Excel has the same problem, but it can be solved during the import process if Excel is forced to treat the digits as text rather than numerical. I don't see a way to do the equivalent in JASP. All other columns import perfectly.

There is already a frequency tab full of useful stats options, however, I really need the KS test option and believe it would be an easier statistic to patch in.

Thank you for all the work, and I like the software. I've spent the last few weeks on Rstat, and this is refreshing. Just need that damn KS test.

Brody

In teaching JASP with undergraduate statistics classes, the one thing that I was completely unable to do with it was Chi Square Goodness of Fit. That would be an excellent addition!

It will be nice to have this feature in JASP. Thanks.

Dear all,

I created a fake set of data to compare the results of a Log-Linear Regression and a Bayesian Log-Linear Regression (see attached .jasp file).

The Log-Linear Regression gives the expected results (IV1 and IV2 p<0.001 and IV1_IV2 p<0.01), but, the Bayesian Log-Linear Regression gives a BF10 of 1. Interestingly, it seems the Bayesian Log-Linear Regression _always* gives a BF10 of 1, independently of the data you feed it.

Cheers!

Log-linear.jasp.zip

Is it possible to add a reporting feature which will convert the statistical tables into words and meanings?
For instance, take the example shown here about ANCOVA : http://www.statisticshell.com/docs/ancova.pdf
You can take the reporting sections out of this file, and create a template that will use specific fields in JASP analysis and embed them in the textual report. If for instance, you request to add an effect size, this effect size will also appear in the textual report of the statistical test.
I THINK IT WILL BE GREAT!

I think it would be nice and very useful in the sequential analysis to not only have a plot of the BF and how it developes over time while the data come in, but also the Kullback-Leibler distance between prior before datapoint N comes in, and the posterior after datapoint N has been processed. This gives a measure of how informative or 'surprising' incoming datapoints are. Normally, this KB measure would be expected to asymptotically approach zero with increasing number of data points, but deviations from this pattern are usually interesting and important.

https://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence

There is no possibility to assign weights to the contrast analysis. Assigning weights is critical to conduct planned comparisons and to ensure the orthogonality of the comparisons. A planned comparison uses the error factor of the omnibus ANOVA and cannot be replaced by any other procedure. For example, in a one-way ANOVA with three levels, a planned comparison should be available for A1 vs. A2 + A3, corresponding to weights 2, -1, -1. Or with 5 levels, a contrast of -3, -3, 2, 2, 2 should be available. In standard programs, the user plugs in the weights, whose sum is 1, corresponding to the required hypothesis.
Again, this is required for both between-subjects designs and within subject designs

Currently, JASP starts with locale set to U.S. My guess this was done as a "fix" for issue #549.
This "breaks" all the locale-sensitive string manipulation done by QT.
The fix is to use QString instead of std::string for manipulating user input and user presentation (and while that is being done, replace std::string for QString everywhere where there is no real need for std::string).
That done we can allow the JASP to use the system locale - giving dd-mm-yyyy format dates and comma as decimal separator, for NL users, dd/mm/yyyy and dot as dec. sep. for UK and mm/dd/yyyy and dot for US.
Putting application locale as an option in the config files (default: 'system') would allow users to choose what is appropriate for them.
Note: For those not familiar with QT locales see: https://wiki.qt.io/Locales

Since measures of reliability (such as Cronbach's alpha) are being added to the next patch, I was thinking that intraclass correlations would be useful for researchers using external rater data or responses coded by judges.

JASP really needs to have more keyboard shortcuts. Suggestions include:

• Keyboard shortcut for each analysis
• Shortcut to add a variable to test column (this is vital, moving mouse to click each variable is time consuming when variables are not ordered the same as your factors).
• Shortcut to switch between different viewers

It is very helpful to have the median absolute deviation as non-parametric description of the dispersion of data.

I noticed that when selecting "Correlated positively" in Bayesian Correlation Pairs test, the Bayes factor in the prior and posterior plots is renamed to BF+0, but the menu buttons are still named BF10, BF01, Log(BF10).

I can imagine that this will confuse some users/students, so you might want to consider renaming the subscripts in the menu as well when you replace "1" with "+" (or "-") in the plots.

Support data in long format, not just wide format (if I'm correct that long format is not supported). Would allow for easy use of R with same dataset. (e.g., my PhD student put data in wide for JASP, then we have to first melt into long to do R analyses)

Strings are inherently ordered, especially if used to represent date and time data, e.g. from a .SAV file. Shouldn't JASP reflect this?

Hi!

I have just been introduced to your software today, and I'm really exited. I am a big supporter of bayesian statistics, and it's great to be able to compare it to the standard methods. So first, thank you for making this technique more accesible!

My request would be to have a feature that does latent class analysis (with bayesian stats). I would love a feature that allows me to see if, for example, I can distinguish different nationalities (classes) based on hight, bodyweight and hair color. What I actually want to do is try latent class analysis on biomarkers in depression. Can I explain the low anti-depressant response rate by showing that there is only one group (class!) of people that has the problem the current medicine is for? Are there other classes that would benefit more from a different biological treatment?

The catch is, I would love to do that with bayesian, but I have no idea how to combine the two. Anyway, I think latent class analysis is very usefull in medical research and a bunch of other fields, and a feature on this amazing program would perhaps help to make it more popular, and (with the inclusion of bayesian) more scientifically sound.

Let me know if you like my idea! If not, I'm still gratefull for the rest!
Lian

In addition to RM-ANOVA, it would be great to also have the ability to run Multilevel (aka. nested, mixed, hierarchical (HLM)) linear models (MLM) and generalized linear model (GzMLM) options (e.g. poisson, binomial, gamma with various link functions: log, logit, etc). I almost never use RM-ANOVA but use MLM and GzMLM all the time in SPSS (as well as SAS and R). Thank you for your great work on JASP so far!! We're seriously considering it as a replacement to SPSS in the future!

As mentioned in the meeting:
An import from database(s) might be useful to some (commercial) users.
First thoughts on construction:

1. Use an ODBC connector to access the databases.
2. Have the user enter an SQL statement ("JOIN") somewhere (An "Import options" dialogue?).
   A full featured DB browser would be overkill.
   Points to think about:

• DB type to JASP type mapping.
• A RDMS could easily produce more data than the memory available to JASP can hold.
• If people are using large datasets (from a DB or elsewhere), might we get performance issues?

Order restrictions are only implemented for the indep. multinomial, and greyed out for the other choices (such as dependent multinomial). This is fine. Now suppose I select "indep. multinomial" and pick the order restriction "group 1 > group 2". This is still fine. But now I change my mind and alter "indep. multinomial" to "dependent multinomial". Then the user interface still indicates that I have an order-restriction, even though this is not implemented.

@FransMeerhoff
@boutinb
Annotation of output is an important feature of JASP. But currently the user cannot change the font size or the font color. It would be great to see that functionality added.

Cheers,
E.J.

Please, add this option:

• Factor / Principal.. = add "save scores" (Bartllet or regression *)
• Factor / Exploratory... = add "save scores" (Bartllet or regression *)
• = or show the factor weights.

@raviselker: it would be nice to add the possibility to correct p-values from multiway ANOVA for multiplicity. See http://arxiv.org/abs/1412.3416 and singmann/afex#3. Another way to add something that SPSS lacks. I think this option should not be default, but available through a checkbox instead.
Cheers, E.J.

Would it be possible to get factor scores saved for principal component and factor analyses?

It would also be pretty interesting to see a Bayesian version of PCA or FA. I've tried figuring out some of the packages in R about Bayesian PCA, and while I am pretty decent with R, the packages available were a bit difficult to use.

When running analyses with a lot of variables (possibly with very long variable names), it would be very useful if the windows within the analysis specifications were rescalable (i.e., get larger if you rescale the size of the global analysis window).

Moreover, when you drag variables into the analysis window, the list of variables always resets to the first variable. This is very exhausting when you want to add 10+ variables to an analysis and have to assign them correctly (e.g., repeated measures-ANOVA). It would be great if the list stayed (optionally?) at the point of the last variable you selected.

From the forum -- we could look into this and when we like it we could offer it as an option or provide it in addition to the current d:
"I know your focus is more on the Bayesian analyses, but a brief comment on the classic t-tests: The Cohen's d reported in the output table for (classic) paired t-tests seems to be uncorrected for the correlation between observations. This can lead to overestimation of the effect size according to some authors (e.g., [1] ) Perhaps it would be good to mention that the d value is uncorrected?
[1] alternative formulae for example in Dunlap, W. P., Cortina, J. M., Vaslow, J. B., & Burke, M. J. (1996). Meta-analysis of experiments with matched groups or repeated measures designs. Psychological Methods, 1(2), 170."

Could you incorporate permutation-based analyses, e.g. for 2-group comparisons of experimental data? Unlike traditional statistics, these involve few assumptions to generate a p value.

It would also be great to build in a permutation-based effect size - for experimental designs, I would recommend one developed by Berry and Mielke (1992) - http://epm.sagepub.com/content/52/1/41.short

Abstract: A new measure of association is introduced which measures the degree of association between a nominal independent variable and nominal, ordinal, or interval dependent variables. An extension to multivariate problems provides analyses of multiple dependent variables, including mixtures of interval, ordinal, and/or nominal dependent variables. A commensuration technique is given for standardizing dependent variables. A permutation test of significance is provided for the new measures.

Hopefully,

Paul

Hi there,

It would be great if, for ANCOVA, in the assumption checks section there was an option to plot comparative regression slopes to allow the visual checking of the homogeneity of regression slopes assumption.

Something similar to the plot shown here for example [(https://goo.gl/images/6wiOSb)]

Thanks

Andy

Environment (for enhancement requests)

• Enhancement: Include credibility intervals in the descriptives for all BF tests
• Purpose: Graphs for BF tests include credibility intervals, but exact values for these are not included in the descriptive tables. Would it be possible to include them in a future release?