{"id":277,"date":"2026-03-31T12:36:53","date_gmt":"2026-03-31T04:36:53","guid":{"rendered":"https:\/\/yoyodyne.com.au\/?p=277"},"modified":"2026-03-31T14:26:42","modified_gmt":"2026-03-31T06:26:42","slug":"ucl-v-utl","status":"publish","type":"post","link":"https:\/\/yoyodyne.com.au\/index.php\/2026\/03\/31\/ucl-v-utl\/","title":{"rendered":"UCL v UTL"},"content":{"rendered":"\n

This dataset $[0.1, 0.2, 0.05]$ is the sample set used in Hewett, 2006 so we’re going to use it here too. It perfectly illustrates why the Upper Tolerance Limit (UTL)<\/strong> (the defensive choice for occupational hygiene), and the Upper Confidence Limit (UCL)<\/strong> (serves a different purpose than ‘most exposed worker in a group’) both fail when sample sizes are small.<\/p>\n\n\n\n

To paraphrase multiple regulators I’ve spoken to:<\/p>\n\n\n\n

“We are not as interested in the average exposure to a work group as the potentially highest exposure end of the workgroup exposure distribution”.<\/p>\n\n\n\n

This is where the langauge gets loose and terms are used interchangably; specfically UCL<\/strong> and UTL<\/strong>.<\/p>\n\n\n\n

The Core Difference<\/h3>\n\n\n\n
    \n
  • UCL (of the Mean):<\/strong> Estimates where the average<\/strong> exposure lies, or the limit below which you are confident it sit under.<\/li>\n<\/ul>\n\n\n\n

    NORM.DIST formula: $mean+ 1.645 \\times sd$
    NORM.DIST formula: $\\bar{x} + 1.645 \\times \\sigma$<\/p>\n\n\n\n

    <\/p>\n\n\n\n

    LOGNORM.DIST: $GM \\times GSD^{1.645}$<\/p>\n\n\n\n

      \n
    • UTL (95\/95):<\/strong> Estimates the limit where the worst-case (95th percentile) of the worse-case of exposures of the population – not the sample set- lie with high confidence.<\/strong><\/li>\n<\/ul>\n\n\n\n
      \n\n\n\n

      The $k$-factor for small datasets<\/h2>\n\n\n\n

      The $k$-factor<\/strong> is a statistical multiplier used to calculate one-sided tolerance limits (like the 95\/95 UTL) when the true population mean and standard deviation are unknown. Unlike a standard Z-score\u2014which assumes you have perfect knowledge of the entire population\u2014the $k$-factor explicitly accounts for the sampling error<\/strong> inherent in small datasets. It acts as a “safety margin” or uncertainty penalty: because we are estimating the distribution from only a few data points, $k$ must be significantly larger than a Z-score to ensure we maintain our desired confidence level (e.g., 95%) that a specific proportion of the population is actually covered. <\/p>\n\n\n\n

      As the sample size ($n$) increases and our estimate of the population becomes more reliable, the $k$-factor decreases, eventually converging toward the standard normal quantile ($1.645$ for the 95th percentile) as $n$ approaches infinity.<\/p>\n\n\n\n

      \n\n\n\n

      Comparison of $k$ vs. $z$ (95% Confidence \/ 95% Coverage)<\/h3>\n\n\n\n
      Sample Size (n)<\/strong><\/td>k-factor<\/strong><\/td>Z-score (z0.95\u200b)<\/strong><\/td>The “Uncertainty Penalty”<\/strong><\/td><\/tr><\/thead>
      3<\/strong><\/td>7.656<\/strong><\/td>1.645<\/td>+365%<\/td><\/tr>
      10<\/strong><\/td>2.911<\/strong><\/td>1.645<\/td>+77%<\/td><\/tr>
      30<\/strong><\/td>2.220<\/strong><\/td>1.645<\/td>+35%<\/td><\/tr>
      $\\infty$<\/strong><\/td>1.645<\/strong><\/td>1.645<\/td>0%<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
      \n\n\n\n

      The Worked Example<\/h2>\n\n\n\n

      With only three data points, the uncertainty is massive. Even though your “best guess” (point estimate) for the 95th percentile is low, the statistical safety margin required for 95% confidence is huge.<\/p>\n\n\n\n

      Metric<\/strong><\/td>Calculation Logic<\/strong><\/td>Result<\/strong><\/td>Context<\/strong><\/td><\/tr><\/thead>
      Point Estimate (95th %ile)<\/strong><\/td>$\\exp^{(\\bar{y} + 1.645 \\cdot s_y)}$<\/td>0.31<\/strong><\/td>“The most likely 95th percentile.”<\/td><\/tr>
      UCL95 (of the Mean)<\/strong><\/td>Land’s H-statistic calculation<\/td>~2.8<\/strong><\/td>“We are 95% sure the average<\/em> is below 2.8.”<\/td><\/tr>
      UTL 95\/95<\/strong><\/td>$\\exp^{(\\bar{y} + 7.656 \\cdot s_y)}$<\/td>20.2<\/strong><\/td>“We are 95% sure 95% of shifts are below 20.2.”<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
      \n\n\n\n

      How to Calculate It<\/h2>\n\n\n\n

      The “jump” from 0.31 to 20.2 happens because of the $k$-factor<\/strong>, which penalizes the small sample size ($n=3$).<\/p>\n\n\n\n

      Log-Transform the data<\/strong><\/p>\n\n\n\n

      $\\ln(0.05) = -3.0$, <\/p>\n\n\n\n

      $\\ln(0.1) = -2.3$, <\/p>\n\n\n\n

      $\\ln(0.2) = -1.6$<\/p>\n\n\n\n

      Descriptive statistics<\/strong><\/p>\n\n\n\n

      $\\bar{y} = -2.3$, $GM = 1.0$<\/p>\n\n\n\n

      $s_y = 0.693$, $GSD = 2.00$<\/p>\n\n\n\n

      Select the K-Factor<\/strong><\/p>\n\n\n\n

      Account for n=3. For a 95\/95 limit with 3 samples, use $k = 7.656$<\/strong>. (For $n=20$, this would drop to ~2.4).<\/p>\n\n\n\n

      Calculate the Limit<\/strong><\/p>\n\n\n\n

      $UTL = \\exp^{(-2.3 + (7.656 \\cdot 0.693))} = \\exp^{(3.0)} \\approx \\mathbf{20.2}$.<\/p>\n\n\n\n

      \n\n\n\n

      The “Sample Size Penalty” Simulator<\/h3>\n\n\n\n

      You can use this tool to see how increasing your sample size ($n$) collapses the massive gap between your point estimate (0.31) and your defensible 95\/95 limit (20.2).<\/p>\n\n\n\n

      \n\n

      The 95\/95 Tolerance Simulator<\/h2>\n \n

      This tool demonstrates the difference between a Point Estimate<\/strong> (a best guess) and the 95\/95 UTL<\/strong> (the defensible compliance limit). As sample size increases, the gap between these two numbers narrows.<\/p>\n\n

      \n \n
      \n
      \n