Totals, Missings, and Denominators

Population Data in the Denominator

What do you do when your target dataset doesn’t have the information necessary to create your denominator? For example - when you create an adverse event table, the adverse event dataset likely only contains records for subjects who experienced an adverse event. But subjects who did not have an adverse event are still part of the study population and must be considered in the denominator.

For this reason, Tplyr allows lets you set a separate population dataset - but there are a couple things you need to do to trigger Tplyr to use the population data as your denominator.

Consider these two examples.

tplyr_table(adae, TRTA) %>% 
  add_layer(
    group_count(AEDECOD) %>% 
      set_distinct_by(USUBJID) %>% 
      set_format_strings(f_str('xx (xx.x%)', distinct_n, distinct_pct))
  ) %>% 
  build() %>% 
  head() %>% 
  kable()

tplyr_table(adae, TRTA) %>% 
  set_pop_data(adsl) %>%
  set_pop_treat_var(TRT01A) %>%
  add_layer(
    group_count(AEDECOD) %>% 
      set_distinct_by(USUBJID) %>% 
      set_format_strings(f_str('xx (xx.x%)', distinct_n, distinct_pct))
  ) %>% 
  build() %>% 
  head() %>% 
  kable()

There are three things done above that use the population data:

• Population data must be set in the first place using set_pop_where() and the population treatment variable must be specified using set_pop_treat_var()
• set_distinct_by() must be used. Why? Because when you need to use a separate population dataset, the target dataset is likely more than one record per subject. So set_distinct_by() in this scenario would be used to get distinct counts per subject. In this example, we’re looking at the number of unique subjects who experienced a specific adverse event.
• The population denominator is then used when calculating distinct_pct. It is also worth noting that the default count layer formats use distinct values.

Note that if you need more control over the values used in the denominator from the population data, you can set a separate filter on the population data used set_pop_where().

Fortunately, denominators are much simpler when they’re kept within a single dataset. Just kidding! Let’s get weird.

Denominator Grouping

When you’re looking within a single dataset, there are a couple factors that you need to consider for a denominator. The first is which grouping variables create those denominators. Let’s look at this from two perspectives - count layers and shift layers.

adsl <- adsl %>% 
  mutate(DCSREAS = ifelse(DCSREAS == '', 'Completed', DCSREAS))
         
tplyr_table(adsl, TRT01P) %>% 
  add_layer(
    group_count(DCSREAS)
  ) %>% 
  build() %>% 
  kable()

tplyr_table(adsl, TRT01P) %>% 
  add_layer(
    group_count(DCSREAS, by=SEX)
  ) %>% 
  build() %>% 
  kable()

tplyr_table(adsl, TRT01P) %>% 
  add_layer(
    group_count(DCSREAS, by=SEX) %>% 
      set_denoms_by(SEX, TRT01P)
  ) %>% 
  build() %>% 
  kable()

tplyr_table(adlb, TRTA, where=PARAMCD == "CK") %>%
  add_layer(
    group_shift(vars(row = BNRIND, column = ANRIND), by = vars(PARAM, AVISIT)) %>%
      set_format_strings(f_str("xx (xxx.x%)", n, pct)) %>%
      # This is the default, the 3x3 box formed by the target variables
      set_denoms_by(TRTA, PARAM, AVISIT) 
  ) %>%
  build() %>%
  kable()

tplyr_table(adlb, TRTA, where=PARAMCD == "CK") %>%
  add_layer(
    group_shift(vars(row = BNRIND, column = ANRIND), by = vars(PARAM, AVISIT)) %>%
      set_format_strings(f_str("xx (xxx.x%)", n, pct)) %>%
      set_denoms_by(TRTA, PARAM, AVISIT, BNRIND) # Each row made by TRTA, BNRIND
  ) %>%
  build() %>%
  arrange(ord_layer_1, ord_layer_2, ord_layer_3) %>% 
  head() %>% 
  kable()

tplyr_table(adlb, TRTA, where = PARAMCD == "CK") %>%
  add_layer(
    group_shift(vars(row = BNRIND, column = ANRIND), by = vars(PARAM, AVISIT)) %>%
      set_format_strings(f_str("xx (xx.xx%)", n, pct)) %>%
      set_denoms_by(TRTA, ANRIND) # Use the column total as the denominator
  ) %>%
  build() %>%
  arrange(ord_layer_1, ord_layer_2, ord_layer_3) %>% 
  head() %>%
  kable()

Controlling the Denominator Filter

There are some circumstances that you’ll encounter where the filter used for a denominator needs to be different than the filter used to count. Disposition tables are an example of this, and we’ll use that example to paint this picture.

‘Tplyr’ offers you the ability to specifically control the filter used within the denominator. This is provided through the function set_denom_where(). The default for set_denom_where() is the layer level where parameter, if one was supplied. set_denom_where() allows you to replace this layer level filter with a custom filter of your choosing. This is done on top of any filtering specified in the tplyr_table() where parameter - which means that the set_denom_where() filter is applied in addition to any table level filtering.

Yeah we know - there are a lot of different places that filtering can happen…

So let’s take the example shown below. The first layer has no layer level filtering applied, so the table level where is the only filter applied. The second layer has a layer level filter applied, so the denominators will be based on that layer level filter. Notice how in this case, the percentages in the second layer add up to 100%. This is because the denominator only includes values used in that layer.

The third layer has a layer level filter applied, but additionally uses set_denom_where(). The set_denom_where() in this example is actually removing the layer level filter for the denominators. This is because in R, when you filter using TRUE, the filter returns all records. So by using TRUE in set_denom_where(), the layer level filter is effectively removed. This causes the denominator to include all values available from the table and not just those selected for that layer - so for this layer, the percentages will not add up to 100%. This is important - this allows the percentages from Layer 3 to sum to the total percentage of “DISCONTINUED” from Layer 1.

adsl2 <- adsl %>% 
  mutate(DISCONTEXT = if_else(DISCONFL == 'Y', 'DISCONTINUED', 'COMPLETED'))

t <- tplyr_table(adsl2, TRT01P, where = SAFFL == 'Y') %>%
  add_layer(
    group_count(DISCONTEXT)
  ) %>%
  add_layer(
    group_count(DCSREAS, where = DISCONFL == 'Y')
  ) %>%
  add_layer(
    group_count(DCSREAS, where = DISCONFL == 'Y') %>% 
    set_denom_where(TRUE)
  ) %>%
  build() %>%
  arrange(ord_layer_index, ord_layer_1) 

t %>% 
  kable()

Missing Counts

Missing counts are a tricky area for frequency tables, and they play directly in with denominators as well. These values raise a number of questions. For example, do you want to format the missing counts the same way as the event counts? Do you want to present missing counts with percentages? Do missing counts belong in the denominator?

The set_missing_count() function can take a new f_str() object to set the display of missing values. If not specified, the associated count layer’s format will be used. Using the ... parameter, you are able to specify the row label desired for missing values and values that you determine to be considered ‘missing’. For example, you may have NA values in the target variable, and then values like “Not Collected” that you also wish to consider “missing”. set_missing_count() allows you to group those together. Actually - you’re able to establish as many different “missing” groups as you want - even though that scenario is fairly unlikely.

In the example below 50 random values are removed and NA is specified as the missing string. This leads us to another parameter - denom_ignore. By default, if you specify missing values they will still be considered within the denominator, but when you have missing counts, you may wish to exclude them from the totals being summarized. By setting denom_ignore to TRUE, your denominators will ignore any groups of missing values that you’ve specified.

adae2 <- adae
adae2[sample(nrow(adae2), 50), "AESEV"] <- NA

t <- tplyr_table(adae2, TRTA) %>%
  add_layer(
    group_count(AESEV) %>%
      set_format_strings(f_str("xxx (xx.xx%)", n, pct)) %>%
      set_missing_count(f_str("xxx", n), sort_value=Inf, denom_ignore=TRUE, Missing = NA)
  ) %>%
  build() %>% 
  arrange(ord_layer_1)

t %>% 
  kable()

We did one more other thing worth explaining in the example above - gave the missing count its own sort value. If you leave this field null, it will simply be the maximum value in the order layer plus 1, to put the Missing counts at the bottom during an ascending sort. But tables can be sorted a lot of different ways, as you’ll see in the sort vignette. So instead of trying to come up with novel ways for you to control where the missing row goes - we decided to just let you specify your own value.

Adding a ‘Total’ Row

In addition to missing counts, some summaries require the addition of a ‘Total’ row. ‘Tplyr’ has the helper function add_total_row() to ease this process for you. Like most other things within ‘Tplyr’ - particularly in this vignette - this too has a significant bit of nuance to it.

Much of this functionality is similar to set_missing_count(). You’re able to specify a different format for the total, but if not specified, the associated count layer’s format will be used. You’re able to set your own sort value to specify where you want the total row to sit.

More nuance comes in two places:

• By default, add_total_row() will count missing values, but you can exclude those values using the count_missings parameter. ‘Tplyr’ will warn you when set_count_missing() has denom_ignore set to TRUE, add_total_row() has count_missings set to TRUE and the format contains a percentage. Why? Because if the denominator is ignoring missing values but you’re still counting them in your total, the percentage shown can exceed 100%.
• add_total_row() will throw a warning when a by variable is used, because it becomes ambiguous what total should be calculated. You can rectify this by using set_denoms_by(), which allows the user to control exactly which groups are used to form the denominator. This way the totals presented by add_total_row() will align with denominators specified in set_denom_by() and generate total rows that match the grouping of your denominator values.

In the example below, we summarize age groups by sex. The denominators are determined by treatment group and sex, and since we are not excluding any values from the denominator, the total row ends up matching the denominator that was used. The ‘Missing’ row tells us the number of missing values, but because count_missings is set to TRUE, the missing counts are included in the total row. This probably isn’t how you would choose to display things, but here we’re trying to show the flexibility built into ‘Tplyr’.

adsl2 <- adsl
adsl2[sample(nrow(adsl2), 50), "AGEGR1"] <- NA

tplyr_table(adsl2, TRT01P) %>% 
  add_layer(
    group_count(AGEGR1, by=SEX) %>% 
      set_denoms_by(TRT01P, SEX) %>%  # This gives me a Total row each group
      add_total_row(f_str("xxx", n), count_missings=TRUE, sort_value=-Inf) %>% 
      set_total_row_label("All Age Groups") %>% 
      set_missing_count(f_str("xx (xx.x%)", n, pct), Missing = NA, sort_value=Inf)
  ) %>% 
  build() %>% 
  arrange(ord_layer_1, ord_layer_2) %>% 
  kable()

The default text for the Total row is “Total”, but we provide set_total_row_label() to allow you to customize the text used in your display.

Let’s look at a more practical version of the table above. If you display missings, you probably want to exclude them from the total. Here we do that using set_missing_count(). So more commonly, you’ll see this:

tplyr_table(adsl2, TRT01P) %>% 
  add_layer(
    group_count(AGEGR1, by=SEX) %>% 
      set_denoms_by(TRT01P, SEX) %>%  # This gives me a Total row each group
      add_total_row(f_str("xxx", n), count_missings=FALSE, sort_value=-Inf) %>% 
      set_total_row_label("All Age Groups") %>% 
      set_missing_count(f_str("xxx", n), Missing = NA, sort_value=Inf, denom_ignore=TRUE)
  ) %>% 
  build() %>% 
  arrange(ord_layer_1, ord_layer_2) %>% 
  kable()

Now the table is more intuitive. We used set_missing_count() to update our denominators, so missing have been excluded. Now, the total row intuitively matches the denominators used within each group, and we can see how many missing records were excluded.

And that’s it for denominators! Happy counting!