ucd-ipo / agroft Goto Github PK

See: #35 (comment)

Do we need to support other operating systems other than Windows?

It's redundant since the power transformation can do this. Also sq root is more common for count data which we aren't allowing for the dependent variable.

This file is extremely difficult to understand. It is a ~800 line function with god knows how many nested if statements. It'll need some major cleanup if any more plotting functionality needs to be added.

The packages shinyBS and readxl are on CRAN now, so the install_github part of the initialize_AIP() function for those packages can be changed to install.packages and the section of code loading readxl and using read_excel can be uncommented. Users should now be able to upload excel files!

See: #34 (comment)

Add functions for multiple mean comparisons that separate treatment means using 3 methods and order them with corresponding letters using Fisher’s protected and unprotected least significant difference (LSD) and Tukey procedures. Add an option to choose significance level.
*Add a reminder that pops up that reminds them that mean separations should only be done if the that factor (treatment) is significant in the ANOVA table (i.e. p-value < their chosen sig level), unless they are doing Fisher’s Unprotected LSD.

@lmpincus Help! I can't find this!

Check for normality and homogeneity of variance using Shapiro Wilk and Levene’s HOV tests, respectively.

• Shapiro-Wilk
• Levene's

The button should be disabled until there is sufficient info.

processing file: report.Rmd
  |.......                                                          |  11%
   inline R code fragments

  |..............                                                   |  22%
label: unnamed-chunk-1 (with options) 
List of 4
 $ echo   : logi FALSE
 $ warning: logi FALSE
 $ error  : logi FALSE
 $ message: logi FALSE

  |......................                                           |  33%
   inline R code fragments

  |.............................                                    |  44%
label: unnamed-chunk-2 (with options) 
List of 3
 $ echo   : logi FALSE
 $ warning: logi FALSE
 $ message: logi FALSE

  |....................................                             |  56%
  ordinary text without R code

  |...........................................                      |  67%
label: unnamed-chunk-3 (with options) 
List of 4
 $ echo   : logi FALSE
 $ results: chr "asis"
 $ warning: logi FALSE
 $ message: logi FALSE

  |...................................................              |  78%
  ordinary text without R code

  |..........................................................       |  89%
label: unnamed-chunk-4 (with options) 
List of 3
 $ echo   : logi FALSE
 $ message: logi FALSE
 $ warning: logi FALSE

  |.................................................................| 100%
   inline R code fragments


output file: report.md

Warning in file.rename(out, file) :
  cannot rename file '_analysis_report_2015-07-21.html' to '/tmp/RtmpXJMAG9/file31004bc0aff0.html', reason 'Invalid cross-device link'
Error opening file: 2
Error reading: 9

• In Analysis tab, under 'Model Fit Summary', add subtitle, 'ANOVA table'.
• Also move this output to below assumption tests.

Right now we are using intxplot from the HH library which requires some C libs to install. We avoid that library altogether by using the builtin interaction.plot.

instead of:
summary(one.df.model)

use:
anova(one.df.model)

It is GPL code and may not be needed. See issue #22 for more details.

It is loaded in global.R, but doesn't seem to be used.

BSD, MIT, or Apache are preferable.

Unless there is a specific reason for needing GPL features, i.e. forcing downstream users to use the same license, then we should remove that restriction and choose a more liberal license that only requires essentially "citation".

From @msimmond:

This is what I was asking about at the last meeting. Jason, do you think there's anyway to link the presentation materials I'm creating in powerpoint to the app in github? I'd like people to be able to use the ppt materials, but with a citation.

Sounds like we need some way to host general materials like a Powerpoint and have them accessible to from the application.

In Analysis output

This could be an option at the end of the LSD step. The function produces LSD table with detransformed means. If this is selected, the post-hoc plots use the detransformed means.

"error: subscript out of bounds"

It still seems to work, but the error pops up.

To reproduce use the corn data with a linear regression:

observation.pow ~ rx

• The user should be able to produce standard bar charts, histograms, and scatterplots that include confidence intervals if relevant.
• The titles, x and y axes, and colors of the plots should be editable via the GUI.
• The interaction plot should be available if model includes an interaction.
• For the linear regression model, the user should be able to generate residuals vs. fitted graph to test assumptions.
• Bar chart comparing the means in ANOVA with standard error for each variable?
• interaction.plot (Maegan will send me an example one)
• Look at residual plots in ANOVA. residuals on y axes and predicted values on x axes (to evaluate whether they are meeting the assumptions of there tests)
• Scatter plot for single variable with R^2 and trend line.
• Plots need to be moved into the analyses steps instead of at the end in step 5.

• Must be able to analyze CRD (completely randomized design) and RCBD, there should be option for 1 and 2 treatments, and their interaction:
  • Y ~ trtA
  • Y ~ trtA trtB trtA*trtB
  • Y ~ trtA Block
  • Y ~ trtA trtB trtA*trtB Block
• Add a reminder that pops up anytime there’s a significant treatment x treatment interaction, that main effects can not be evaluated.

Two things:

• Can you please update the parameters for the plot(model...) so that there is less white space between subplots? I found this to work:
  par(mfrow = c(2, 1), oma = c(0, 0, 2, 0),
  mar = c(4,4,2,2))

-Can you please change the 2 plots produced to: (1) res x pred (already there) and (2) histogram or kernel density plot (I think we were calling it) - instead of QQ

• shinyjs can do this but it requires R > 3.1
• Here is a propsoal by the shinyjs author to add reset functionality to shiny: rstudio/shiny#870

The user should be able to select the location as a random effect and the mixed effect model results computed. Should include a pop-up box explaining why location is being analyzed as a random effect.

Can you remove all rows in the Tukey's ANOVA table except for 'sq.pred'? If not, add subtitle to 'Tukey's Test for Nonadditivity': 'Attention: refer only to 'sq.pred' in this table - ignore all other rows'

To reproduce, use runApp('inst/app') to start the server and the refresh the browser window. The server disconnects, R unblocks, and there are no error messages.

I found a shiny app with a similar idea as this one.

https://vnijs.shinyapps.io/base/?SSUID=2e1f1c7227d5dc9bbff1b3a311982f0c

Check it out. Website here:

http://vnijs.github.io/radiant/

Change the squared predicted variable name to something consistent for any scenario, such as 'sq.pred'.

The report is useful, but an R script with all the code needed to reproduce the GUI analyses might be even better. This is not in the deliverables of the contract.

Following up on #13 here's an example of a variance-weighted ANOVA that can be performed if transformations don't resolve the problems with assumption tests for the scenario 'CRD with 1 treatment':

Instead of assuming equal variance with: model <- aov(y ~ group, data = my.data)
Do variance-weighted ANOVA that does not assume equal variances. There could be option below transformation drop-down to do that, but don't allow transformation + variance-weight ANOVA.
oneway.test(y ~ group, data = my.data)
(Source: https://stat.ethz.ch/R-manual/R-devel/library/stats/html/oneway.test.html)

So Lauren said we should call this the "Agricultural Field Trial Statistics Package". That is a mouthful so we need something simpler for the library name. We could make an acronym: "AFTSP" or "AFTS" but I'm not really fond of that. Ian named it "AIP", which I don't even know what it stands for.

What do you all think about "Chaval"? It is the Urdu word for "rice".

So "Chaval: An Agricultural Field Trial Statistics Package" and you would install and load it with:

install.packages('chaval')
library('chaval')
chaval()

Currently, we are doing the following: if there is a significant interaction between 2 treatment factors (A and B) in a CRD or RCBD fixed effects model (e.g. y ~ A + B + A:B + Blk), we perform an LSD on the combination of all levels of both factors (e.g., in a 3 x 3 factorial, we would do an LSD on 9 treatments), as opposed to first subsetting the data for each level of A and re-running the ANOVA on B (y ~ B + Blk #for each level of A) and doing LSD test if significant effect of B, and vice versa for A. I am not sure if both approaches are valid (i.e. looking at all combinations versus subsetting). In the latter approach we'd have to retest all ANOVA assumptions for each individual ANOVA. YUCK

This is what Ian used:

http://oddhypothesis.blogspot.com/2014/04/deploying-self-contained-r-apps-to.html

There should be a dropdown function under ANOVA for split plot methods. R function available in the agricolae package (split plot only at this time).

The app should be able to print a PDF report of the functions performed.

Add a popup over the graph images that remind people that they can save the images by right clicking.

Transforms: log, square root, power. Add a pop-up that suggests/explains different types of transformations and when they are used.

From the console:

> library('agricolae')
> data('plots')
> spcrd.fit <- aov(yield ~ A + B + A:B + Error(A:block), plots)
> summary(spcrd.fit)

Error: A:block
          Df Sum Sq Mean Sq F value Pr(>F)
A          1 0.1071  0.1071   0.143   0.77
Residuals  1 0.7500  0.7500               

Error: Within
          Df Sum Sq Mean Sq F value   Pr(>F)    
A          1   4.20    4.20   1.172   0.3045    
B          2  29.78   14.89   4.155   0.0486 *  
A:B        2 300.44  150.22  41.922 1.37e-05 ***
Residuals 10  35.83    3.58                     

---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
> sprcbd.fit <- aov(yield ~ A + B + A:B + block + Error(A:block), plots)
> summary(sprcbd.fit)

Error: A:block
      Df Sum Sq Mean Sq
A      1 0.1071  0.1071
block  1 0.7500  0.7500

Error: Within
          Df Sum Sq Mean Sq F value   Pr(>F)    
A          1   4.20    4.20   1.172   0.3045    
B          2  29.78   14.89   4.155   0.0486 *  
A:B        2 300.44  150.22  41.922 1.37e-05 ***
Residuals 10  35.83    3.58                     

---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

And results from the app for spcrd:

And for sprcbd:

Error: A:block
          Df Sum Sq Mean Sq F value Pr(>F)
A          1  0.222  0.2222   0.108  0.774
block      2  2.111  1.0556   0.514  0.661
Residuals  2  4.111  2.0556               

Error: Within
          Df Sum Sq Mean Sq F value   Pr(>F)    
B          2  29.78   14.89   3.458 0.082744 .  
A:B        2 300.44  150.22  34.890 0.000112 ***
Residuals  8  34.44    4.31                     

---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Add USAID branding and project logo: could just have a link to a website or a window pop-up under “about” for app.

We'll have to figure out how to do this, but for now as a reminder...
Excerpt from "http://www.unh.edu/halelab/BIOL933/lectures/lect_16_reading.pdf":
'In summary, a major problem in the analysis of unbalanced data is the contamination of means,
and thus the differences among means, by effects of other factors. The solution to these
problems is to replace missing data by their least square estimates and to remove the
contaminating effects of other factors through a proper adjustment of means. With R, all this
means is that you should use the Type II SS (Anova()) and the lsmeans() function.'

• Remove t-test as the same thing can be accomplished with one-way ANOVA of two factors.
• We need a selection box for experiment design that comes before analyses type. This should have a selection for CRD/RCBD/None and a checkbox to signal whether this is a split plot design if CRD or RCBD is selected or not. The subsequent steps in the analysis side panel will be adjusted (or automatically selected) based on this selection. The variable choices for split plot should reflect the main plot and split variable. The split plot may use aov or lm depending on what type of design it is (see the example in #15).
• The analyses type should be ANOVA or linear model or selected automatically based on the split plot selections.
• Remove RCBD from analysis, it is an experiment design type.
• Maybe only do single variate linear model.

See the design in the wiki: https://github.com/ucd-ipo/aip-analysis/wiki

I'd like to move this application to a Github organization so it isn't under Kyle or my username. @msimmond, what would be a good org name? Is this under your lab? Or is it under a project?

Add disclaimer (e.g. to best of our knowledge this is correct and we cannot be held accountable for any conclusions or data generated using this)