computing differences between groups: alternative to spread for multiple computations - r

Question

I commonly need to compute differences between groups, nested by some interval and/or additional grouping. For computing a single variable, this is easy to accomplish with spread and mutate. Here's a reproducible example with the datasetChickWeight; don't get distracted by the calculation itself (this is just a toy example), my question is about how to handle a dataset structured like the dataframe ChickSum created below.
# reproducible dataset
data(ChickWeight)
ChickSum = ChickWeight %>%
filter(Time == max(Time) | Time == min(Time)) %>%
group_by(Diet, Time) %>%
summarize(mean.weight = mean(weight)) %>%
ungroup()
Here is how I might go about calculating the change in average chick weight between the first and last time, stratified by diet:
# Compute change in mean weight between first and last time
ChickSum %>%
spread(Time, mean.weight) %>%
mutate(weight.change = `21` - `0`)
However, this doesn't work so well with multiple variables:
ChickSum2 = ChickWeight %>%
filter(Time == max(Time) | Time == min(Time)) %>%
group_by(Diet, Time) %>%
# now also compute variable "count"
summarize(count = n(), mean.weight = mean(weight)) %>%
ungroup()
I can't spread by Time and both count and mean.weight; my current solution is to do two spread-mutate operations---once for count and again for mean.weight---and then join the results.
ChickCountChange = ChickSum2 %>%
select(-mean.weight) %>%
spread(Time, count) %>%
mutate(count.change = `21` - `0`)
ChickWeightChange = ChickSum2 %>%
select(-count) %>%
spread(Time, mean.weight) %>%
mutate(weight.change = `21` - `0`)
full_join(
select(ChickWeightChange, Diet, weight.change),
select(ChickCountChange, Diet, count.change),
by = "Diet")
Is there another approach to these types of computation? I've been trying to conceive of a strategy that combines group_by and purrr::pmap in order to avoid spread but still maintain the advantages of the above approach (such as spread's fill argument for choosing how to handle missing group combinations), but I haven't figured it out. I'm open to suggestions or alternative data structures/ways of thinking about the problem.

Answer 1

You might try re-grouping, then using lag() to calculate the differences. Works for your toy example, but it may be better to see some of your real dataset:
ChickWeight %>%
filter(Time == max(Time) | Time == min(Time)) %>%
group_by(Diet, Time) %>%
# now also compute variable "count"
summarize(count = n(), mean.weight = mean(weight)) %>%
ungroup() %>%
group_by(Diet) %>%
mutate(count.change = count - lag(count),
weight.change = mean.weight - lag(mean.weight)) %>%
filter(Time == max(Time))
Result:
Diet Time count mean.weight count.change weight.change
<fct> <dbl> <int> <dbl> <int> <dbl>
1 1 21 16 178. -4 136.
2 2 21 10 215. 0 174
3 3 21 10 270. 0 230.
4 4 21 9 239. -1 198.

Answer 2

So I came up with a potential/partial solution in the process of writing up a reproducible example. Essentially, we use gather to group by the variables themselves:
ChickSum2 %>%
gather(variable, value, count, mean.weight) %>%
spread(Time, value) %>% mutate(Change = `21` - `0`) %>%
select(Diet, variable, Change) %>%
spread(variable, Change)
This works only if the following two conditions are true:
All variables are the same type (e.g. both mean.weight and count are numeric).
the difference calculation is the same for all variables (e.g. I want to compute last - first for all variables).
I guess the second condition could be relaxed by using e.g. case_when.