Example

The following code estimates optimal lambda and SEz for nine centiles spaced two-thirds of a z-score apart, based on a sample of 10,000 children.

knitr::kable(optimal_design(z = -4:4*2/3, N = 10000), digits = c(2, 2, 0, 3, 0, 2, 2))

What the table shows is unsurprising, that a sample of 10,000 children estimates the median (p = 50th centile) much more precisely than the 0.4th centile (and by symmetry the 99.6th centile), with SEz 0.033 compared to 0.065. The 95% confidence interval for the 50th centile is (47.3, 52.7) and that for the 0.4th centile is (0.26, 0.56).

In addition lambda increases from 0.38 on the median to 0.90 on the 0.4th and 99.6th centiles, indicating a large shift in sample composition. Let’s see what the corresponding age distributions look like.

minage <- 0
maxage <- 20
N <- 1e4
map_dfr(c(1, 0.9, 0.38), ~{
  tibble(age = (runif(N, minage^.x, maxage^.x))^(1/.x),
         p = .x)
}) %>% 
  mutate(lambda = fct_reorder(factor(p), -p))  %>%
  ggplot(aes(age, group = lambda)) +
  geom_histogram(fill = 'gray', binwidth = 1) +
  facet_wrap(. ~ lambda, labeller = label_both) +
  xlab('age (years)') + ylab('frequency')

The figure shows the impact of lambda on the age distribution. When lambda = 1 sampling is uniform across the age range – the histogram shows one-year age groups with about 500 per group. In practice this design is rarely appropriate for growth studies.

With lambda = 0.9, suitable for the 0.4th and 99.6th centiles, the sampling is close to uniform.

However with lambda = 0.38, which is appropriate for estimating the median, there is dramatic oversampling in early infancy. This is needed to accurately estimate the shape of the median curve soon after birth when it is at its steepest.

Example 1

The following code calculates the age group sizes optimised for centiles from the 0.4th to the 99.6th, with a sample of 10,000 children from 0 to 20 years in one-year groups.

  n_table <- map_dfc(-4:4*2/3, ~{
    n_agegp(z = .x, N = 10000) %>% 
      select(!!z2cent(.x) := n_varying)
  }) %>% 
  bind_cols(tibble(age = paste(0:19, 1:20, sep = '-')), .)
knitr::kable(n_table)

The table shows how the sample composition varies depending on which centile is optimised. Comparing the 50th and 0.4th centiles, the optimal number for the 50th centile is more than four times the size in the first year but less than half the size at age 19-20. This matches the histogram above.

Example 2

n_agegp can also be used to design studies over a narrower age range, e.g. 0-5 years. For example there are 3506 children in this age range in the table optimised for the 2nd centile, and the following code returns the same group sizes (bar rounding):

knitr::kable(n_agegp(z = 2, N = 3506, minage = 0, maxage = 5, n_groups = 5) %>% 
  select(age, n_varying))

This shows that the sample composition optimised for age 0-20 is also optimal for subsets of the age range.

Example 3

A third example contrasts the Cuban Growth Study design, as mentioned earlier, with the optimal design based on the specified precision of SEz = 0.053 on the 3rd centile, which Healy (1974) showed required 1000 boys aged 8.

For an optimal design the required precision should apply across the age range. The code shows the age distribution for 0-20 years that was used in the Study, including a 10% uplift in numbers to give 1100 at age 8, and the corresponding design had it been optimised.

# define CGS age distribution
  tibble(age = c(1/6, 1/2, 5/6, 5/4, 7/4, 19/8, 25/8, 4:10,
                 c(22:29/2 - 1/4), 15:18, 77/4),
         n = c(rep(12, 3), rep(10, 4), 13, rep(11, 6), rep(6, 3),
               rep(11, 5), 15, rep(8, 3), 13) * 100,
         span = c(rep(1/3, 3), rep(1/2, 2), rep(3/4, 2), rep(1, 7),
                  rep(1/2, 8), rep(1, 4), 3/2)) %>% 
    ggplot(aes(x = age, y = n/span, width = span-0.02)) +
    xlab('age (years)') + ylab('frequency') +
    geom_bar(fill = 'gray', stat = 'identity') +
    geom_bar(aes(y = n_varying, width = NULL), 
             data = n_agegp(z = qnorm(0.03), SEz = 0.3 / 5.75) %>% 
               mutate(n_varying = n_varying * 1.1), 
             width = 0.98, fill = 'gray50', stat = 'identity')

The pale gray histogram shows the complex sample composition used in the Cuban Growth Study, with oversampling in infancy and puberty reflecting the higher growth velocity at those ages.

However the pubertal oversampling is unnecessary, and the design ignores the saving in numbers that is achieved by smoothing the centile curves.

The darker histogram shows the optimal design. Overall the sample size could have been reduced from 28,000 to around 11,000. This highlights the saving in resources that optimisation can achieve.