If you find any typos, errors, or places where the text may be improved, please let us know
by providing feedback either in the feedback survey (given during class) or
by using GitHub.
On GitHub open an
issue
or submit a pull request
by clicking the " Edit this page" link at the side of this page.
8Efficiently running many analyses at once
Instructor note
Before beginning, get them to recall what they remember of the previous session, either with something like Mentimeter or verbally. Preferably something like Mentimeter because it allows everyone to participate, not just the ones who are more comfortable being vocal to the whole group.
Depending on what they write, might need to briefly go over the previous session.
Rarely do we run only one single statistical model to answer one single question, especially in our data-overflowing environments. An initial instinct when faced with this task might be to copy-and-paste, then slightly modify the code each time. Or, if you have heard of loops or used them in other programming languages, you might think to create a loop. Thankfully R uses something more powerful and expressive than either of those approaches, and that is functional programming. Using functional programming concepts, we can use little code to express complex actions and run large numbers of statistical analyses.
Connected to the concept of functional programming, is the idea of resampling a dataset multiple times and running the statistical analysis on each resampled set to calculate a more accurate measure of uncertainty. We often use generic calculations of uncertainty like the standard error or the confidence interval. Those are useful measures especially with very large datasets, however, they have limitations of their own. By making use of resampling, we can identify how uncertain or unreliable a statistical result might be for our specific dataset. This session will be about using functional programming in the context of statistical analysis and learning about other methods of determining uncertainty.
8.1 Learning objectives
The overall objective for this session is to:
Describe the basic framework underlying most statistical analyses and use R to generate statistical results using this framework.
More specific objectives are to:
Recall principles of functional programming and apply them to running statistical analyses by using the {purrr} package.
Describe what a resampling technique is, the types available, and why it can help estimate the variability of model results. Apply functions from {rsample} and {tune} to use these techniques.
Continue applying the concepts and functions used from the previous sessions.
8.2 Exercise: How would we use functional programming to run multiple models?
Time: ~20 minutes.
Functional programming underlies many core features of running statistical methods on data. This exercise is meant for you to review this concept and try to think of it in the context of statistical modeling.
For 8 minutes, discuss with your neighbour how we can use functional programming to apply the statistical model to each metabolite. Try to think how the code would look. You don’t need to write real R code, but if writing pseudocode helps, go right ahead! Also, don’t look ahead 😉
For the remaining time, we will discuss our thoughts in the whole group.
Instructor note
After they’ve finished, either write pseudocode in RStudio or draw this out on a whiteboard if it is available. There will probably be several different approaches, many of which could also be implemented just fine. Ultimately we will replace create_recipe_spec(metabolite_...) with create_recipe_spec(starts_with("metabolite_")).
8.3 Apply logistic regression to each metabolite
You may have thought of many different ways to run the model on each metabolite based on the lipidomics_wide dataset. However, these types of “split-apply-combine” tasks are (usually) best done using data in the long form. So we’ll start with the original lipidomics dataset. Create a header and code chunk at the end of the doc/lesson.qmd file:
## Running multiple models```{r}```
The first thing we want to do is convert the metabolite names into snake case:
# A tibble: 504 × 6
code gender age class metabolite value
<chr> <chr> <dbl> <chr> <chr> <dbl>
1 ERI109 M 25 CT tms_interntal_standard 208.
2 ERI109 M 25 CT cholesterol 19.8
3 ERI109 M 25 CT lipid_ch_3_1 44.1
4 ERI109 M 25 CT lipid_ch_3_2 147.
5 ERI109 M 25 CT cholesterol 27.2
6 ERI109 M 25 CT lipid_ch_2 587.
7 ERI109 M 25 CT fa_ch_2_ch_2_coo 31.6
8 ERI109 M 25 CT pufa 29.0
9 ERI109 M 25 CT phosphatidylethanolamine 6.78
10 ERI109 M 25 CT phosphatidycholine 41.7
# ℹ 494 more rows
The next step is to split the data up. We could use group_by(), but in order to make the most use of {purrr} functions like map(), we will use group_split() to convert the data frame into a set of lists1. Let’s first add {purrr} as a dependency:
use_package("purrr")
Then we run group_split() on the metabolite column, which will output a lot of data frames as a list.
<list_of<
tbl_df<
code : character
gender : character
age : double
class : character
metabolite: character
value : double
>
>[12]>
[[1]]
# A tibble: 36 × 6
code gender age class metabolite value
<chr> <chr> <dbl> <chr> <chr> <dbl>
1 ERI109 M 25 CT cd_cl_3_solvent 166.
2 ERI111 M 39 CT cd_cl_3_solvent 171.
3 ERI163 W 58 CT cd_cl_3_solvent 262.
4 ERI375 M 24 CT cd_cl_3_solvent 172.
5 ERI376 M 26 CT cd_cl_3_solvent 300.
6 ERI391 M 31 CT cd_cl_3_solvent 241.
7 ERI392 M 24 CT cd_cl_3_solvent 172.
8 ERI79 W 26 CT cd_cl_3_solvent 148.
9 ERI81 M 52 CT cd_cl_3_solvent 168.
10 ERI83 M 25 CT cd_cl_3_solvent 253.
# ℹ 26 more rows
[[2]]
# A tibble: 108 × 6
code gender age class metabolite value
<chr> <chr> <dbl> <chr> <chr> <dbl>
1 ERI109 M 25 CT cholesterol 19.8
2 ERI109 M 25 CT cholesterol 27.2
3 ERI109 M 25 CT cholesterol 8.88
4 ERI111 M 39 CT cholesterol 22.8
5 ERI111 M 39 CT cholesterol 30.2
6 ERI111 M 39 CT cholesterol 9.28
7 ERI163 W 58 CT cholesterol 14.9
8 ERI163 W 58 CT cholesterol 24.0
9 ERI163 W 58 CT cholesterol 7.66
10 ERI375 M 24 CT cholesterol 19.2
# ℹ 98 more rows
[[3]]
# A tibble: 36 × 6
code gender age class metabolite value
<chr> <chr> <dbl> <chr> <chr> <dbl>
1 ERI109 M 25 CT fa_ch_2_ch_2_coo 31.6
2 ERI111 M 39 CT fa_ch_2_ch_2_coo 28.9
3 ERI163 W 58 CT fa_ch_2_ch_2_coo 36.6
4 ERI375 M 24 CT fa_ch_2_ch_2_coo 39.4
5 ERI376 M 26 CT fa_ch_2_ch_2_coo 52.1
6 ERI391 M 31 CT fa_ch_2_ch_2_coo 42.8
7 ERI392 M 24 CT fa_ch_2_ch_2_coo 39.9
8 ERI79 W 26 CT fa_ch_2_ch_2_coo 32.7
9 ERI81 M 52 CT fa_ch_2_ch_2_coo 28.4
10 ERI83 M 25 CT fa_ch_2_ch_2_coo 26.5
# ℹ 26 more rows
[[4]]
# A tibble: 36 × 6
code gender age class metabolite value
<chr> <chr> <dbl> <chr> <chr> <dbl>
1 ERI109 M 25 CT lipid_ch_2 587.
2 ERI111 M 39 CT lipid_ch_2 585.
3 ERI163 W 58 CT lipid_ch_2 558.
4 ERI375 M 24 CT lipid_ch_2 606.
5 ERI376 M 26 CT lipid_ch_2 554.
6 ERI391 M 31 CT lipid_ch_2 597.
7 ERI392 M 24 CT lipid_ch_2 607.
8 ERI79 W 26 CT lipid_ch_2 546.
9 ERI81 M 52 CT lipid_ch_2 593.
10 ERI83 M 25 CT lipid_ch_2 606.
# ℹ 26 more rows
[[5]]
# A tibble: 36 × 6
code gender age class metabolite value
<chr> <chr> <dbl> <chr> <chr> <dbl>
1 ERI109 M 25 CT lipid_ch_3_1 44.1
2 ERI111 M 39 CT lipid_ch_3_1 28.1
3 ERI163 W 58 CT lipid_ch_3_1 75.1
4 ERI375 M 24 CT lipid_ch_3_1 22.0
5 ERI376 M 26 CT lipid_ch_3_1 29.5
6 ERI391 M 31 CT lipid_ch_3_1 38.0
7 ERI392 M 24 CT lipid_ch_3_1 34.8
8 ERI79 W 26 CT lipid_ch_3_1 109.
9 ERI81 M 52 CT lipid_ch_3_1 49.6
10 ERI83 M 25 CT lipid_ch_3_1 29.9
# ℹ 26 more rows
[[6]]
# A tibble: 36 × 6
code gender age class metabolite value
<chr> <chr> <dbl> <chr> <chr> <dbl>
1 ERI109 M 25 CT lipid_ch_3_2 147.
2 ERI111 M 39 CT lipid_ch_3_2 153.
3 ERI163 W 58 CT lipid_ch_3_2 144.
4 ERI375 M 24 CT lipid_ch_3_2 220.
5 ERI376 M 26 CT lipid_ch_3_2 282.
6 ERI391 M 31 CT lipid_ch_3_2 220.
7 ERI392 M 24 CT lipid_ch_3_2 215.
8 ERI79 W 26 CT lipid_ch_3_2 153.
9 ERI81 M 52 CT lipid_ch_3_2 150.
10 ERI83 M 25 CT lipid_ch_3_2 153.
# ℹ 26 more rows
[[7]]
# A tibble: 36 × 6
code gender age class metabolite value
<chr> <chr> <dbl> <chr> <chr> <dbl>
1 ERI109 M 25 CT mufa_pufa 50.6
2 ERI111 M 39 CT mufa_pufa 53.2
3 ERI163 W 58 CT mufa_pufa 60.7
4 ERI375 M 24 CT mufa_pufa 0.532
5 ERI376 M 26 CT mufa_pufa 1.15
6 ERI391 M 31 CT mufa_pufa 0.602
7 ERI392 M 24 CT mufa_pufa 0.422
8 ERI79 W 26 CT mufa_pufa 36.3
9 ERI81 M 52 CT mufa_pufa 40.1
10 ERI83 M 25 CT mufa_pufa 39.3
# ℹ 26 more rows
[[8]]
# A tibble: 36 × 6
code gender age class metabolite value
<chr> <chr> <dbl> <chr> <chr> <dbl>
1 ERI109 M 25 CT phosphatidycholine 41.7
2 ERI111 M 39 CT phosphatidycholine 52.9
3 ERI163 W 58 CT phosphatidycholine 35.3
4 ERI375 M 24 CT phosphatidycholine 66.9
5 ERI376 M 26 CT phosphatidycholine 32.7
6 ERI391 M 31 CT phosphatidycholine 62.9
7 ERI392 M 24 CT phosphatidycholine 64.3
8 ERI79 W 26 CT phosphatidycholine 41.0
9 ERI81 M 52 CT phosphatidycholine 56.1
10 ERI83 M 25 CT phosphatidycholine 57.8
# ℹ 26 more rows
[[9]]
# A tibble: 36 × 6
code gender age class metabolite value
<chr> <chr> <dbl> <chr> <chr> <dbl>
1 ERI109 M 25 CT phosphatidylethanolamine 6.78
2 ERI111 M 39 CT phosphatidylethanolamine 3.66
3 ERI163 W 58 CT phosphatidylethanolamine 3.59
4 ERI375 M 24 CT phosphatidylethanolamine 3.59
5 ERI376 M 26 CT phosphatidylethanolamine 2.33
6 ERI391 M 31 CT phosphatidylethanolamine 1.46
7 ERI392 M 24 CT phosphatidylethanolamine 2.00
8 ERI79 W 26 CT phosphatidylethanolamine 4.93
9 ERI81 M 52 CT phosphatidylethanolamine 5.20
10 ERI83 M 25 CT phosphatidylethanolamine 5.01
# ℹ 26 more rows
[[10]]
# A tibble: 36 × 6
code gender age class metabolite value
<chr> <chr> <dbl> <chr> <chr> <dbl>
1 ERI109 M 25 CT phospholipids 5.58
2 ERI111 M 39 CT phospholipids 6.16
3 ERI163 W 58 CT phospholipids 5.19
4 ERI375 M 24 CT phospholipids 4.20
5 ERI376 M 26 CT phospholipids 3.27
6 ERI391 M 31 CT phospholipids 4.71
7 ERI392 M 24 CT phospholipids 4.14
8 ERI79 W 26 CT phospholipids 5.70
9 ERI81 M 52 CT phospholipids 5.46
10 ERI83 M 25 CT phospholipids 4.89
# ℹ 26 more rows
[[11]]
# A tibble: 36 × 6
code gender age class metabolite value
<chr> <chr> <dbl> <chr> <chr> <dbl>
1 ERI109 M 25 CT pufa 29.0
2 ERI111 M 39 CT pufa 27.4
3 ERI163 W 58 CT pufa 35.5
4 ERI375 M 24 CT pufa 6.92
5 ERI376 M 26 CT pufa 3.22
6 ERI391 M 31 CT pufa 3.43
7 ERI392 M 24 CT pufa 3.52
8 ERI79 W 26 CT pufa 18.7
9 ERI81 M 52 CT pufa 20.7
10 ERI83 M 25 CT pufa 18.2
# ℹ 26 more rows
[[12]]
# A tibble: 36 × 6
code gender age class metabolite value
<chr> <chr> <dbl> <chr> <chr> <dbl>
1 ERI109 M 25 CT tms_interntal_standard 208.
2 ERI111 M 39 CT tms_interntal_standard 219.
3 ERI163 W 58 CT tms_interntal_standard 57.1
4 ERI375 M 24 CT tms_interntal_standard 19.2
5 ERI376 M 26 CT tms_interntal_standard 35.4
6 ERI391 M 31 CT tms_interntal_standard 30.4
7 ERI392 M 24 CT tms_interntal_standard 21.7
8 ERI79 W 26 CT tms_interntal_standard 185.
9 ERI81 M 52 CT tms_interntal_standard 207.
10 ERI83 M 25 CT tms_interntal_standard 322.
# ℹ 26 more rows
Remember that logistic regression models need each row to be a single person, so we’ll use the functional map() to apply our custom function metabolites_to_wider() on each of the split list items:
[[1]]
# A tibble: 36 × 5
code gender age class metabolite_cd_cl_3_solvent
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 166.
2 ERI111 M 39 CT 171.
3 ERI163 W 58 CT 262.
4 ERI375 M 24 CT 172.
5 ERI376 M 26 CT 300.
6 ERI391 M 31 CT 241.
7 ERI392 M 24 CT 172.
8 ERI79 W 26 CT 148.
9 ERI81 M 52 CT 168.
10 ERI83 M 25 CT 253.
# ℹ 26 more rows
[[2]]
# A tibble: 36 × 5
code gender age class metabolite_cholesterol
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 18.6
2 ERI111 M 39 CT 20.8
3 ERI163 W 58 CT 15.5
4 ERI375 M 24 CT 10.2
5 ERI376 M 26 CT 13.5
6 ERI391 M 31 CT 9.53
7 ERI392 M 24 CT 9.87
8 ERI79 W 26 CT 17.6
9 ERI81 M 52 CT 17.0
10 ERI83 M 25 CT 19.7
# ℹ 26 more rows
[[3]]
# A tibble: 36 × 5
code gender age class metabolite_fa_ch_2_ch_2_coo
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 31.6
2 ERI111 M 39 CT 28.9
3 ERI163 W 58 CT 36.6
4 ERI375 M 24 CT 39.4
5 ERI376 M 26 CT 52.1
6 ERI391 M 31 CT 42.8
7 ERI392 M 24 CT 39.9
8 ERI79 W 26 CT 32.7
9 ERI81 M 52 CT 28.4
10 ERI83 M 25 CT 26.5
# ℹ 26 more rows
[[4]]
# A tibble: 36 × 5
code gender age class metabolite_lipid_ch_2
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 587.
2 ERI111 M 39 CT 585.
3 ERI163 W 58 CT 558.
4 ERI375 M 24 CT 606.
5 ERI376 M 26 CT 554.
6 ERI391 M 31 CT 597.
7 ERI392 M 24 CT 607.
8 ERI79 W 26 CT 546.
9 ERI81 M 52 CT 593.
10 ERI83 M 25 CT 606.
# ℹ 26 more rows
[[5]]
# A tibble: 36 × 5
code gender age class metabolite_lipid_ch_3_1
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 44.1
2 ERI111 M 39 CT 28.1
3 ERI163 W 58 CT 75.1
4 ERI375 M 24 CT 22.0
5 ERI376 M 26 CT 29.5
6 ERI391 M 31 CT 38.0
7 ERI392 M 24 CT 34.8
8 ERI79 W 26 CT 109.
9 ERI81 M 52 CT 49.6
10 ERI83 M 25 CT 29.9
# ℹ 26 more rows
[[6]]
# A tibble: 36 × 5
code gender age class metabolite_lipid_ch_3_2
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 147.
2 ERI111 M 39 CT 153.
3 ERI163 W 58 CT 144.
4 ERI375 M 24 CT 220.
5 ERI376 M 26 CT 282.
6 ERI391 M 31 CT 220.
7 ERI392 M 24 CT 215.
8 ERI79 W 26 CT 153.
9 ERI81 M 52 CT 150.
10 ERI83 M 25 CT 153.
# ℹ 26 more rows
[[7]]
# A tibble: 36 × 5
code gender age class metabolite_mufa_pufa
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 50.6
2 ERI111 M 39 CT 53.2
3 ERI163 W 58 CT 60.7
4 ERI375 M 24 CT 0.532
5 ERI376 M 26 CT 1.15
6 ERI391 M 31 CT 0.602
7 ERI392 M 24 CT 0.422
8 ERI79 W 26 CT 36.3
9 ERI81 M 52 CT 40.1
10 ERI83 M 25 CT 39.3
# ℹ 26 more rows
[[8]]
# A tibble: 36 × 5
code gender age class metabolite_phosphatidycholine
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 41.7
2 ERI111 M 39 CT 52.9
3 ERI163 W 58 CT 35.3
4 ERI375 M 24 CT 66.9
5 ERI376 M 26 CT 32.7
6 ERI391 M 31 CT 62.9
7 ERI392 M 24 CT 64.3
8 ERI79 W 26 CT 41.0
9 ERI81 M 52 CT 56.1
10 ERI83 M 25 CT 57.8
# ℹ 26 more rows
[[9]]
# A tibble: 36 × 5
code gender age class metabolite_phosphatidylethanolamine
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 6.78
2 ERI111 M 39 CT 3.66
3 ERI163 W 58 CT 3.59
4 ERI375 M 24 CT 3.59
5 ERI376 M 26 CT 2.33
6 ERI391 M 31 CT 1.46
7 ERI392 M 24 CT 2.00
8 ERI79 W 26 CT 4.93
9 ERI81 M 52 CT 5.20
10 ERI83 M 25 CT 5.01
# ℹ 26 more rows
[[10]]
# A tibble: 36 × 5
code gender age class metabolite_phospholipids
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 5.58
2 ERI111 M 39 CT 6.16
3 ERI163 W 58 CT 5.19
4 ERI375 M 24 CT 4.20
5 ERI376 M 26 CT 3.27
6 ERI391 M 31 CT 4.71
7 ERI392 M 24 CT 4.14
8 ERI79 W 26 CT 5.70
9 ERI81 M 52 CT 5.46
10 ERI83 M 25 CT 4.89
# ℹ 26 more rows
[[11]]
# A tibble: 36 × 5
code gender age class metabolite_pufa
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 29.0
2 ERI111 M 39 CT 27.4
3 ERI163 W 58 CT 35.5
4 ERI375 M 24 CT 6.92
5 ERI376 M 26 CT 3.22
6 ERI391 M 31 CT 3.43
7 ERI392 M 24 CT 3.52
8 ERI79 W 26 CT 18.7
9 ERI81 M 52 CT 20.7
10 ERI83 M 25 CT 18.2
# ℹ 26 more rows
[[12]]
# A tibble: 36 × 5
code gender age class metabolite_tms_interntal_standard
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 208.
2 ERI111 M 39 CT 219.
3 ERI163 W 58 CT 57.1
4 ERI375 M 24 CT 19.2
5 ERI376 M 26 CT 35.4
6 ERI391 M 31 CT 30.4
7 ERI392 M 24 CT 21.7
8 ERI79 W 26 CT 185.
9 ERI81 M 52 CT 207.
10 ERI83 M 25 CT 322.
# ℹ 26 more rows
Alright, we now a list of data frames where each data frame has only one of the metabolites. These bits of code represent the conceptual action of “splitting the data into a list by metabolites”. Since we’re following a function-oriented workflow, let’s create a function for this. Convert it into a function, add the Roxygen comments (have the cursor inside the function, type Ctrl-Shift-P, then type “roxygen”), style using the Palette (, then type “style file”), move into the R/functions.R file, and then source() the file.
#' Convert the long form dataset into a list of wide form data frames.#'#' @param data The lipidomics dataset.#'#' @return A list of data frames.#'split_by_metabolite <-function(data) { data %>%column_values_to_snakecase(metabolite) %>% dplyr::group_split(metabolite) %>% purrr::map(metabolites_to_wider)}
In the doc/lesson.qmd, use the new function in the code:
lipidomics %>%split_by_metabolite()
[[1]]
# A tibble: 36 × 5
code gender age class metabolite_cd_cl_3_solvent
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 166.
2 ERI111 M 39 CT 171.
3 ERI163 W 58 CT 262.
4 ERI375 M 24 CT 172.
5 ERI376 M 26 CT 300.
6 ERI391 M 31 CT 241.
7 ERI392 M 24 CT 172.
8 ERI79 W 26 CT 148.
9 ERI81 M 52 CT 168.
10 ERI83 M 25 CT 253.
# ℹ 26 more rows
[[2]]
# A tibble: 36 × 5
code gender age class metabolite_cholesterol
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 18.6
2 ERI111 M 39 CT 20.8
3 ERI163 W 58 CT 15.5
4 ERI375 M 24 CT 10.2
5 ERI376 M 26 CT 13.5
6 ERI391 M 31 CT 9.53
7 ERI392 M 24 CT 9.87
8 ERI79 W 26 CT 17.6
9 ERI81 M 52 CT 17.0
10 ERI83 M 25 CT 19.7
# ℹ 26 more rows
[[3]]
# A tibble: 36 × 5
code gender age class metabolite_fa_ch_2_ch_2_coo
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 31.6
2 ERI111 M 39 CT 28.9
3 ERI163 W 58 CT 36.6
4 ERI375 M 24 CT 39.4
5 ERI376 M 26 CT 52.1
6 ERI391 M 31 CT 42.8
7 ERI392 M 24 CT 39.9
8 ERI79 W 26 CT 32.7
9 ERI81 M 52 CT 28.4
10 ERI83 M 25 CT 26.5
# ℹ 26 more rows
[[4]]
# A tibble: 36 × 5
code gender age class metabolite_lipid_ch_2
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 587.
2 ERI111 M 39 CT 585.
3 ERI163 W 58 CT 558.
4 ERI375 M 24 CT 606.
5 ERI376 M 26 CT 554.
6 ERI391 M 31 CT 597.
7 ERI392 M 24 CT 607.
8 ERI79 W 26 CT 546.
9 ERI81 M 52 CT 593.
10 ERI83 M 25 CT 606.
# ℹ 26 more rows
[[5]]
# A tibble: 36 × 5
code gender age class metabolite_lipid_ch_3_1
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 44.1
2 ERI111 M 39 CT 28.1
3 ERI163 W 58 CT 75.1
4 ERI375 M 24 CT 22.0
5 ERI376 M 26 CT 29.5
6 ERI391 M 31 CT 38.0
7 ERI392 M 24 CT 34.8
8 ERI79 W 26 CT 109.
9 ERI81 M 52 CT 49.6
10 ERI83 M 25 CT 29.9
# ℹ 26 more rows
[[6]]
# A tibble: 36 × 5
code gender age class metabolite_lipid_ch_3_2
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 147.
2 ERI111 M 39 CT 153.
3 ERI163 W 58 CT 144.
4 ERI375 M 24 CT 220.
5 ERI376 M 26 CT 282.
6 ERI391 M 31 CT 220.
7 ERI392 M 24 CT 215.
8 ERI79 W 26 CT 153.
9 ERI81 M 52 CT 150.
10 ERI83 M 25 CT 153.
# ℹ 26 more rows
[[7]]
# A tibble: 36 × 5
code gender age class metabolite_mufa_pufa
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 50.6
2 ERI111 M 39 CT 53.2
3 ERI163 W 58 CT 60.7
4 ERI375 M 24 CT 0.532
5 ERI376 M 26 CT 1.15
6 ERI391 M 31 CT 0.602
7 ERI392 M 24 CT 0.422
8 ERI79 W 26 CT 36.3
9 ERI81 M 52 CT 40.1
10 ERI83 M 25 CT 39.3
# ℹ 26 more rows
[[8]]
# A tibble: 36 × 5
code gender age class metabolite_phosphatidycholine
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 41.7
2 ERI111 M 39 CT 52.9
3 ERI163 W 58 CT 35.3
4 ERI375 M 24 CT 66.9
5 ERI376 M 26 CT 32.7
6 ERI391 M 31 CT 62.9
7 ERI392 M 24 CT 64.3
8 ERI79 W 26 CT 41.0
9 ERI81 M 52 CT 56.1
10 ERI83 M 25 CT 57.8
# ℹ 26 more rows
[[9]]
# A tibble: 36 × 5
code gender age class metabolite_phosphatidylethanolamine
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 6.78
2 ERI111 M 39 CT 3.66
3 ERI163 W 58 CT 3.59
4 ERI375 M 24 CT 3.59
5 ERI376 M 26 CT 2.33
6 ERI391 M 31 CT 1.46
7 ERI392 M 24 CT 2.00
8 ERI79 W 26 CT 4.93
9 ERI81 M 52 CT 5.20
10 ERI83 M 25 CT 5.01
# ℹ 26 more rows
[[10]]
# A tibble: 36 × 5
code gender age class metabolite_phospholipids
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 5.58
2 ERI111 M 39 CT 6.16
3 ERI163 W 58 CT 5.19
4 ERI375 M 24 CT 4.20
5 ERI376 M 26 CT 3.27
6 ERI391 M 31 CT 4.71
7 ERI392 M 24 CT 4.14
8 ERI79 W 26 CT 5.70
9 ERI81 M 52 CT 5.46
10 ERI83 M 25 CT 4.89
# ℹ 26 more rows
[[11]]
# A tibble: 36 × 5
code gender age class metabolite_pufa
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 29.0
2 ERI111 M 39 CT 27.4
3 ERI163 W 58 CT 35.5
4 ERI375 M 24 CT 6.92
5 ERI376 M 26 CT 3.22
6 ERI391 M 31 CT 3.43
7 ERI392 M 24 CT 3.52
8 ERI79 W 26 CT 18.7
9 ERI81 M 52 CT 20.7
10 ERI83 M 25 CT 18.2
# ℹ 26 more rows
[[12]]
# A tibble: 36 × 5
code gender age class metabolite_tms_interntal_standard
<chr> <chr> <dbl> <chr> <dbl>
1 ERI109 M 25 CT 208.
2 ERI111 M 39 CT 219.
3 ERI163 W 58 CT 57.1
4 ERI375 M 24 CT 19.2
5 ERI376 M 26 CT 35.4
6 ERI391 M 31 CT 30.4
7 ERI392 M 24 CT 21.7
8 ERI79 W 26 CT 185.
9 ERI81 M 52 CT 207.
10 ERI83 M 25 CT 322.
# ℹ 26 more rows
Like we did with the metabolite_to_wider(), we need to pipe the output into another map() function that has a custom function running the models. We don’t have this function yet, so we need to create it. Let’s convert the modeling code we used in the exercise above into a function, replacing lipidomics with data and using starts_with("metabolite_") within the create_recipe_spec(). Add the Roxygen comments (have the cursor inside the function, type Ctrl-Shift-P, then type “roxygen”), use the Palette (, then type “style file”) to style, move into the R/functions.R file, and then source() the file.
#' Generate the results of a model#'#' @param data The lipidomics dataset.#'#' @return A data frame.#'generate_model_results <-function(data) {create_model_workflow( parsnip::logistic_reg() %>% parsnip::set_engine("glm"), data %>%create_recipe_spec(tidyselect::starts_with("metabolite_")) ) %>% parsnip::fit(data) %>%tidy_model_output()}
Then we add it to the end of the pipe, but using map_dfr() to convert to a data frame:
Since we are only interested in the model results for the metabolites, let’s keep only the term rows that are metabolites using filter() and str_detect().
Wow! We’re basically at our first {targets} output! Before continuing, there is one aesthetic thing we can add: The original variable names, rather than the snake case version. Since the original variable still exists in our lipidomics dataset, we can join it to the model_estimates object with right_join(), along with a few other minor changes. First, we’ll select() only the metabolite and then create a duplicate column of metabolite called term (to match the model_estimates) using mutate().
Awesome 😄 Now can you guess what we are going to do next? That’s right, making a function of both the model creation code and this code to add the original variable names. Then we can add our first {targets} output!
8.4 Exercise: Creating functions for model results and adding as a target in the pipeline
Time: ~25 minutes.
Convert the code that calculates the model estimates as well as the code that adds the original metabolite names into functions. Start with the code for the metabolite names, using the scaffold below as a starting point.
Name the new function add_original_metabolite_names.
Within the function(), add two arguments, where the first is called model_results and the second is called data.
Paste the code we created into the function, replacing lipidomics with data and model_estimates with model_results.
Add dplyr:: and stringr:: before their respective functions.
Add the Roxygen comments (have the cursor inside the function, type Ctrl-Shift-P, then type “roxygen”).
Use the Palette (, then type “style file”) to style the file to fix up the code.
Cut and paste the function over into the R/functions.R file.
Commit the changes you’ve made so far.
___ <-function(___, ___) { ___ %>%}
Click for the solution. Only click if you are struggling or are out of time.
#' Add the original metabolite names (not as snakecase) to the model results.#'#' @param model_results The data frame with the model results.#' @param data The original lipidomics dataset.#'#' @return A data frame.#'add_original_metabolite_names <-function(model_results, data) { data %>% dplyr::mutate(term = metabolite) %>%column_values_to_snakecase(term) %>% dplyr::mutate(term = stringr::str_c("metabolite_", term)) %>% dplyr::distinct(term, metabolite) %>% dplyr::right_join(model_results, by ="term")}
Do the same thing with the code that creates the model results, using the scaffold below as a starting point.
calculate_estimates <-function(data) { ___ %>%# All the other code to create the results ___ %>%add_original_metabolite_names(data) }
Click for the solution. Only click if you are struggling or are out of time.
#' Calculate the estimates for the model for each metabolite.#'#' @param data The lipidomics dataset.#'#' @return A data frame.#'calculate_estimates <-function(data) { data %>%column_values_to_snakecase(metabolite) %>% dplyr::group_split(metabolite) %>% purrr::map(metabolites_to_wider) %>% purrr::map_dfr(generate_model_results) %>% dplyr::filter(stringr::str_detect(term, "metabolite_")) %>%add_original_metabolite_names(data)}
Lastly, add the model results output to end of the _targets.R file, using the below scaffold as a guide.
Use df_model_estimates for the name.
Use the calculate_estimates() function in command, with lipidomics as the argument.
Use the Palette (, then type “style file”) to style and than run targets::tar_visnetwork() (Ctrl-Shift-P, then type “targets visual”) to see if the new target gets detected. If it does, than run targets::tar_make() (Ctrl-Shift-P, then type “targets run”).
Commit the changes to the Git history.
list( ...,list(name = ___,command =___(___) ))
8.5 Visualizing the model estimates
We’ve got one target done for the modeling stage, three more to go! There are multiple ways of visualizing the results from models. A common approach is to use a “dot-and-whisker” plot like you might see in a meta-analysis. Often the “whisker” part is the measure of uncertainty like the confidence interval, and the “dot” is the estimate. For the confidence interval, we haven’t calculated them at this point because the typical approach doesn’t exactly work for our data (tested before the course). The next section will be covering another way of determining uncertainty. For this plot though, we will use the standard error of the estimate.
Inside the doc/report.Rmd, let’s create a new header and code chunk inside the ## Results section. We’ll want to use tar_read(df_model_estimates) so that {targets} is aware that the R Markdown file is dependent on this target.
### Figure of model estimates```{r}model_estimates <-tar_read(df_model_estimates)```
Then we’ll start using {ggplot2} to visualize the results. For dot-whisker plots, the “geom” we would use is called geom_pointrange(). It requires four values:
x: This will be the “dot”, representing the estimate column.
y: This is the categorical variable that the “dot” is associated with, in this case, it is the metabolite column.
xmin: This is the lower end of the “whisker”. Since the std.error is a value representing uncertainty of the estimate on either side of it (+ or -), we will need to subtract std.error from the estimate.
xmax: This is the upper end of the “whisker”. Like xmin above, but adding std.error instead.
Woah, there is definitely something wrong here. The values of the estimate should realistically be somewhere between 0 (can’t have a negative odds) and 2 (in biology and health research, odds ratios are rarely above 2), definitely unlikely to be more than 5. We will eventually need to troubleshoot this issue, but for now, let’s restrict the x axis to be between 0 and 5.
plot_estimates +coord_fixed(xlim =c(0, 5))
There are so many things we could start investigating based on these results in order to fix them up. But for now, this will do.
8.6 Exercise: Add plot function as a target in the pipeline
Time: ~15 minutes.
Hopefully you’ve gotten comfortable with the function-oriented workflow, because we’ll need to convert this plot code into a function and add it as a target in the pipeline. Use the scaffold below as a guide.
Replace model_estimates with results.
Move the function into the R/functions.R file, add Roxygen comments (have the cursor inside the function, type Ctrl-Shift-P, then type “roxygen”), and use the Palette (, then type “style file”) to style.
Click for the solution. Only click if you are struggling or are out of time.
#' Plot the estimates and standard errors of the model results.#'#' @param results The model estimate results.#'#' @return A ggplot2 figure.#'plot_estimates <-function(results) { results %>% ggplot2::ggplot(ggplot2::aes(x = estimate, y = metabolite,xmin = estimate - std.error,xmax = estimate + std.error )) + ggplot2::geom_pointrange() + ggplot2::coord_fixed(xlim =c(0, 5))}
Then, after doing that, add the new function as a target in the pipeline, name the new name as fig_model_estimates.
And replace all the plot code in the doc/report.Rmd file with the tar_read():
```{{r}}tar_read(fig_model_estimates)```
Run targets::tar_make() (Ctrl-Shift-P, then type “targets run”) to update the pipeline. Then commit the changes to the Git history.
8.7 Determine variability in model estimates with resampling
Depending on the type of research questions, there are several ways to assess variability (or uncertainty) in the model results. We could use the calculated standard error of the estimate or calculate the confidence interval from the standard error (using tidy(conf.int = TRUE)). The disadvantage of this approach is that it isn’t very accurate for the data. Plus, when we have such small sample sizes, some issues can limit the use of these typical measures of uncertainty. And we’ve already noticed that there is something strange with the estimates for some of the metabolites.
So instead, we can use something that is a bit more targeted to the data called “resampling”. There are many resampling techniques, and they all have slightly different uses. The one we will use is called the “bootstrap”2. What bootstrapping does is take the data and randomly resamples it as many times as there are rows. This means you can potentially resample the same data (in our case, person) more than once (called “with replacement”). So by pure chance, you could theoretically have a “resampled set” from lipidomics where all 36 rows are only duplicate data on one person!
What’s the advantage of this? It is a way of directly calculating the standard error from the data itself (rather than from a formula), so it gives a more accurate view of how uncertain the model estimate is for our data. Usually, creating between 50 to 100 “resampled sets” is sufficient to calculate a value for the variation. Because running models with bootstrapped sets can take a long time to process, we will only resample 10 times, or less if your computer is slow.
We’re using the {rsamples} package to handle “resampling”. So let’s add it to our dependency list:
use_package("rsamples")
We will eventually run bootstraps on all the metabolites, so we will need to use our split_by_metabolite() function first. For now, we will only use the first item in that list (accessed with [[1]]) to show that the code works without running on all the metabolites every time. Create another code chunk at the bottom of doc/lesson.qmd to add this code:
You don’t need to run the code in this reading section.
The bootstraps() function is how we create resampled sets. Since this is done randomly, we should use set.seed() in order for the analysis to be reproducible. Nothing is truly random in computers, and instead is actually “pseudorandom”. In order for our analysis to be reproducible, we use set.seed() to force a specific “pseudorandom” value.
set.seed(1324)bootstraps(lipidomics_list[[1]], times =10)
This output is called a “nested tibble”. A nested tibble is a tibble/data frame where one or more of the columns are actually a list object. In our case, each bootstrapped set (marked by the id) has instructions on how the resampled data will look. We can see what it looks like by accessing the splits column and looking at the first item with [[1]]:
bootstraps(lipidomics_list[[1]], times =10)$splits[[1]]
<Analysis/Assess/Total>
<36/15/36>
The contents of this resampled set are split into “analysis” sets and “assessment” sets. You don’t need to worry about what these mean or how to use them, since a function we will later use handles it for us. But to give you an idea of what bootstrapping is doing here, we can access one of the sets with either the analysis() or assessment() functions. We’ll arrange() by code to show how we can have duplicate persons when resampling:
bootstraps(lipidomics_list[[1]], times =10)$splits[[1]] %>%analysis() %>%arrange(code)
# A tibble: 36 × 5
code gender age class metabolite_cd_cl_3_solvent
<chr> <chr> <dbl> <chr> <dbl>
1 ERI140 M 25 T1D 84.6
2 ERI140 M 25 T1D 84.6
3 ERI142 M 38 T1D 70.8
4 ERI143 M 43 T1D 204.
5 ERI144 M 35 T1D 108.
6 ERI144 M 35 T1D 108.
7 ERI144 M 35 T1D 108.
8 ERI145 M 36 T1D 133.
9 ERI145 M 36 T1D 133.
10 ERI145 M 36 T1D 133.
# ℹ 26 more rows
See how some code IDs are the same? Those are the same person that has been selected randomly into this resampled set.
Like we did with the previous modeling, we need to create a workflow object. We’ll use the first metabolite (lipidomics_list[[1]]) for now, but will revise the code to eventually run the bootstrapping on all metabolites.
Previously, we used fit() on the workflow and on the data. Instead, we will use fit_resamples() to run the model on the bootstrapped data. Instead of the data argument in fit(), it is the resamples argument where we provide the bootstraps() sets. We could run the code with only the workflow object and the resampled data, but there’s an extra argument in fit_resamples() that controls some actions taken during the fitting by using the control_resamples() function. For instance, we can save the predictions with save_pred = TRUE and we can process the output with our tidy_model_output() function in the extract argument. So let’s do that. First, both fit_resamples() and control_resamples() come from the {tune} package, so let’s add it to the dependencies first.
use_package("tune")
Now, we can write the code for fit_resamples() on the bootstraps() of the first item in the lipidomics_list and setting the control options with control_resamples().
You’ll see that it gives another nested tibble, but with more columns included. Before we start selecting the results that we want, let’s convert the code above into a function, using the function-oriented workflow we’ve used throughout the course.
#' Generate the model variation results using bootstrap on a single metabolite.#'#' @param data The lipidomics data.#'#' @return A nested tibble.#'generate_model_variation <-function(data) {create_model_workflow( parsnip::logistic_reg() %>% parsnip::set_engine("glm"), data %>%create_recipe_spec(tidyselect::starts_with("metabolite_")) ) %>% tune::fit_resamples(resamples = rsample::bootstraps(data, times =10),control = tune::control_resamples(extract = tidy_model_output,save_pred =TRUE ) )}
Re-writing the code to use the function, it becomes:
Let’s explain this output a bit. The fit_resamples() function outputs a nested tibble, where each row is a resampled set. The columns that begin with . (.metrics or .extracts) are extracted details from each model fit to the resampled set. We’ll ignore all but the .extracts column, since that is the column that we set to extract the tidy_model_output(). Let’s select() only the id and the .extracts column and use unnest() to convert the nested tibble to a regular tibble based on the column given.
Alright, this is actually another nested tibble (we can see based on the new column .extracts where each row is called a <tibble>). So let’s again unnest() this new .extracts column.
Now this is something we are familiar with looking at! It shows the term, the estimate, as well as the bootstrap id. Like before, we only want the metabolite estimate, so we can use filter() and str_detect() like last time, as well as add the original variable names with add_original_metabolite_names().
Using the same workflow as before, let’s convert this into a function:
#' Tidy up the bootstrap output.#'#' @param bootstrap_results The bootstrap object with model results.#'#' @return A data frame.#'tidy_bootstrap_output <-function(bootstrap_results) { bootstrap_results %>% dplyr::select(id, .extracts) %>%# Need to unnest twice since first `.extracts` is a nest of another two# columns of `.extracts` and `.config`. tidyr::unnest(cols = .extracts) %>% tidyr::unnest(cols = .extracts) %>% dplyr::filter(stringr::str_detect(term, "metabolite_")) %>%add_original_metabolite_names(lipidomics)}
Then we can start from the beginning again, right from lipidomics, to split_by_metabolite(), to map()’ing with generate_model_variation(), and finally to map_dfr() with tidy_bootstrap_output(). Keep in mind, this will take a while to run.
8.8 Exercise: Convert to function and add as a target in the pipeline
Time: ~15 minutes.
Continue the workflow we’ve applied throughout the course:
Move the code into a function structure (use the scaffold below as a guide).
Include one argument in the function() called data.
Replace lipidomics in the code with data.
Add the Roxygen comments (have the cursor inside the function, type Ctrl-Shift-P, then type “roxygen”).
Cut and paste the function over into the R/functions.R file.
Style using the Palette (, then type “style file”).
Commit the changes to the Git history.
Use this code as a guide for the function.
calculate_variation <-function(___) { ___ %>%# Code from above. ___}
Click for the solution. Only click if you are struggling or are out of time.
#' Calculate the uncertainty in results.#'#' @param data The lipidomics data.#'#' @return A data frame (or file path)#'calculate_variation <-function(data) { data %>%split_by_metabolite() %>% purrr::map(generate_model_variation) %>% purrr::map_dfr(tidy_bootstrap_output)}
Next, add the function to _targets.R.
Create another tar_target() item in the list() at the bottom of the file.
Use df_model_variation as the name and calculate_variation() as the command with lipidomics as argument.
Run targets::tar_visnetwork() (Ctrl-Shift-P, then type “targets visual”) to see what targets are outdated and then run targets::tar_make() (Ctrl-Shift-P, then type “targets run”).
Commit the changes to the Git history.
Use this code as a scaffold:
list( ...,tar_target(name = ___,command = ___ ))
Click for the solution. Only click if you are struggling or are out of time.
Take your time explaining why we might use a figure like this, and what you’re trying to show. Since there isn’t much code involved, there is time to explain.
Like we did with the estimates, let’s visualize the results. Visualizing the estimates was pretty easy, visualizing the variation is even easier. We want to show the range of estimate values across all the bootstrapped models, by metabolite variable. There’s a neat geom in {ggplot2} called geom_dotplot() that is similar to a histogram, but instead shows individual data points instead of bars. And since we want to show the variation by metabolite, we can use facet_wrap(). We will use scales = "free" because the range of values for estimate are different for each metabolite.
Bin width defaults to 1/30 of the range of the data. Pick better
value with `binwidth`.
This nicely shows the ranges of values in the estimate, really highlighting how uncertain the results are for answering our original research questions. This figure could also be improved quite a bit from a visual and aesthetic point of view, but at least from a content point of view, it shows what we want. For now, we’ll stick with this and finish our last pipeline target before putting everything into the doc/report.Rmd file.
Let’s use our function workflow with this code:
Create the code as a function, add a data argument, and replace the input data object name with data.
Move the code into the R/functions.R file.
Add the Roxygen comments (have the cursor inside the function, type Ctrl-Shift-P, then type “roxygen”) and fill it in.
Style the code using the Palette (, then type “style file”).
#' Plot the uncertainty in the estimates of the models.#'#' @param model_results The model results with the variation.#'#' @return A ggplot2 image.#'plot_variation <-function(model_results) { model_results %>% ggplot2::ggplot(ggplot2::aes(x = estimate)) + ggplot2::geom_dotplot() + ggplot2::facet_wrap(ggplot2::vars(metabolite), scales ="free")}
Then we’ll add it as a pipeline target to the _targets.R file.
Run targets::tar_visnetwork() (Ctrl-Shift-P, then type “targets visual”) and then run targets::tar_make() (Ctrl-Shift-P, then type “targets run”). Once everything has been built, commit everything to the Git history.
8.10 Combine all the output into the R Markdown file
Now its’ time to add the model results and plots to the doc/report.Rmd file. Open it up and create another code chunk at the bottom of the file. Like we did with the other outputs (like the figures), we’ll use tar_read() to reference the image path.
```{{r}}tar_read(fig_model_estimates)```
```{r}tar_read(fig_model_variations)```
Run targets::tar_visnetwork() (Ctrl-Shift-P, then type “targets visual”), then targets::tar_make() (Ctrl-Shift-P, then type “targets run”). We now have the report rendered to an HTML file! If you open it up in a browser, we can see the figures added to it. In the next session we will get more into making the document output nicer looking and creating it as a website.
:)
8.11 Summary
Use functional programming with map(), as part of the function-oriented workflow, to run multiple models efficiently and with minimal code.
Consistently create small functions that do a specific conceptual action and chain them together into larger conceptual actions, which can then more easily be incorporated into a {targets} pipeline. Small, multiple functions combined together are easier to manage than fewer, bigger ones.
Use dot-whisker plots like geom_pointrange() to visualize the estimates and their standard error.
Use resampling techniques like bootstraps(), combined with fit_resamples(), to calculate measures of variation specific to the data. Combine with functionals like map() to run large numbers of models, easily and with minimal code.
Visualize variation in data with alternatives to histograms like geom_dotplot().
There is probably a more computationally efficient way of coding this instead of making a list, but as the saying goes “premature optimization is the root of all evil”. For our purposes, this is a very good approach, but for very large datasets and hundreds of potential models to run, this method would need to be optimized some more.↩︎
Another common technique is called “v-fold cross-validation”, which provides a way of assessing how well the model as a whole performs at fitting the data, rather than bootstrap which determines how varied the estimate can be.↩︎
