Concept map / Lognormal response times
Share
Lognormal response times
Compare the same simulated item response times in seconds and in log-seconds, see why the mean sits above the median, and read the share of responses past a time limit.
1Put the same response times on two scales
Response times on a test item are positive and skewed to the right. Most test takers answer within a typical range, and a few take much longer. A lognormal model says that the natural log of each time follows a normal distribution, with mean μ and standard deviation σ.
If the median response takes 24 seconds, does the average response take 24 seconds too, and how many test takers run past a 60-second limit? Switch between seconds and log-seconds. The large dot is the same simulated response in both views.
With a median of 24 seconds and σ = 0.65 on the log scale, the mean response time is 29.65 seconds, 23.5% above the median. Of all responses, 7.9% take longer than the 60-second limit. In seconds the long right tail pulls the mean above the median, and the 100 simulated responses have sample skewness 2.73. The large dot took 29.17 seconds.
2Average the logs: what comes back?
You record 400 response times to one item. They follow a lognormal model with a median of 24 seconds and σ = 0.65. You take the natural log of each time, average the logs, and turn the average back into seconds with ex. What does that number estimate?
3Normal on the log scale, skewed in seconds
The model says ln T is normal with mean μ and standard deviation σ. Back in seconds, eμ is the median and the mean is larger by the factor eσ²/2. Point at a term and the matching line or area in section 1 lights up.
Now: μ = ln 24 = 3.178, and E[T] = 24 × e0.65²/2 = 24 × 1.235 = 29.65 seconds.
Now: z = (ln 60 − 3.178) / 0.65 = 1.41, so P(T > 60) = 0.079. An interval in seconds and the matching interval in log-seconds hold the same probability. Only the density heights change.
4Check yourself
The tasks use the sliders in section 1 and tick themselves when you meet them.
- Make the mean response time at least 50% larger than the median.Now: σ = 0.65, mean 23.5% above the median
- With σ at least 0.80, set a time limit that no more than 10% of responses exceed.Now: σ = 0.65, 7.9% exceed 60 s
- Make the response times nearly symmetric in seconds: get the mean within 5% of the median.Now: mean 23.5% above the median
5Simulate response times in R
rlnorm() takes meanlog and sdlog, here μ = ln 24 and σ = 0.65. The snippet simulates 1,000 response times, and taking logs makes them symmetric: the skewness drops from 2.47 in seconds to −0.09 in log-seconds. The sample mean, 28.76 s, sits above the median, 23.60 s, while the back-transformed average log, 23.28 s, stays near the median. The model gives a mean of 29.645 s and 0.079 of responses over 60 s.
Simulate lognormal response timesShow the code
# Response times with median 24 s and sigma 0.65 on the
# log scale, the page's defaults: meanlog = log(24)
set.seed(24)
rt <- rlnorm(1000, meanlog = log(24), sdlog = 0.65)
# Moment skewness: 0 for a symmetric distribution
skew <- function(x) {
d <- x - mean(x)
mean(d^3) / mean(d^2)^1.5
}
round(c(seconds = skew(rt), log_seconds = skew(log(rt))), 2)
# The mean sits above the median; exp(mean(log)) does not
round(c(mean = mean(rt), median = median(rt),
back = exp(mean(log(rt)))), 2)
# Model values: mean exp(mu + sigma^2 / 2), P(T > 60)
over_60 <- plnorm(60, log(24), 0.65, lower.tail = FALSE)
model_mean <- exp(log(24) + 0.65^2 / 2)
round(c(mean = model_mean, over_60 = over_60), 3)Uses base R only.
Seconds and log-seconds, in brief
- If ln T is normal with mean μ and standard deviation σ, the time T is lognormal. Taking logs turns right-skewed times into a symmetric set.
- Logs keep the order of the times, so the median is eμ. The mean, eμ+σ²/2, is larger by the factor eσ²/2: 29.6 s against a median of 24 s at σ = 0.65.
- Averaging log-times and transforming back estimates the median, not the mean. A report that does this should say which one it gives.
- The share past a limit L is 1 − Φ((ln L − μ)/σ), 7.9% for a 60-second limit at the default settings. With the median fixed, a larger σ pushes more responses past any limit above the median.
- Van der Linden’s lognormal response-time model gives each test taker a speed and each item a time intensity, with ln T = item intensity − person speed + normal error. It is used to flag rapid guessing and to set time limits, while item response theory models the accuracy of the same responses.
Recommended reading
- e-Handbook of Statistical MethodsNIST/SEMATECH§1.3.6.6.9, the lognormal distribution
- A Lognormal Model for Response Times on Test ItemsWim J. van der LindenPaper · 2006, a psychometric model for item response times