A two-year-old at the 90th percentile for height is more likely than not to be tall as an adult, but the relationship is loose. Cole and Wright (2011) modeled adult height from a child's current height percentile and reported a prediction standard error of about 4 to 5 cm from age 4 through puberty. At age 2 the error is larger.
A percentile at 2 indicates direction. It is a poor estimate of the final number.
The child height predictor does not accept children under 4 for its primary estimate for this reason, and it reports a current percentile without projecting it to adulthood.
Why the correlation is imperfect
Height at 2 reflects the child's growth so far and the conditions of infancy. Adult height additionally reflects growth over the next 14 to 16 years, including the timing and size of the pubertal growth spurt, which is not visible at 2. The correlation between height at 2 and adult height is real, but the unexplained part is large.
Regression to the mean adds to this. Cole and Wright's model adjusts for it. A child at an extreme percentile tends to end up at a less extreme adult percentile. Part of the extreme position at any single age is due to factors that do not persist. A child at the 97th percentile at age 2 has an expected adult percentile below the 97th, and a child at the 3rd has an expected adult percentile above the 3rd.
Percentile crossing in early childhood is also common. Mei and colleagues (2004) found that shifts across two major percentile lines were frequent before age 2 and uncommon afterward. A percentile measured at 24 months may itself still be settling.
What happens when a percentile is projected
The simplest adult-height method reads the child's percentile off the chart at age 2. It then reads the adult height for the same percentile at age 18 or 20 on the CDC 2000 reference. The method assumes no regression to the mean and no change in channel, and it produces an estimate whose error is larger than Cole and Wright's adjusted method.
The predictor shows the current percentile so that a user can see where the child sits. It labels the percentile as a description of today, not an estimate of adulthood.
What improves the estimate
Adding information about the child and the family narrows the error. The corrected Khamis-Roche equation uses exact age, current height, current weight, and the average of the parents' heights. Khamis and Roche (1994) reported average 90% absolute-error bounds of about 5.3 cm for boys and 4.3 cm for girls across ages 4 to 17.5. The equation was fitted to a single, narrow cohort, and the predictor says so on the result page. It is still the most widely used method that does not need an X-ray. It is the predictor's primary estimate from age 4.
Below age 4, the predictor reports the current percentile and the family range from the parents' heights, and states that no adult-height estimate is provided. The family range, about 10 cm either side of the mid-parental target, is wider than the Khamis-Roche bound and is the appropriate level of precision for a child that young.
Timing of measurement
A single measurement at age 2 is the least informative point at which to ask about adult height. It is too late for birth size to add much and too early for the child's own growth pattern to have settled. Between age 4 and the start of puberty, a child's percentile is most stable and the predictor's error is smallest. During puberty, the error rises again because children of the same age differ by years in maturation.
For a question about adult height in a child under 4, the most useful action is to record the measurement and revisit the predictor at 4 or later. For a question about whether a two-year-old's growth is normal, the relevant comparison is with the child's own earlier measurements and the parents' heights. That assessment belongs with the child's pediatrician.