3.4 ppm is the Defect Rate for 4.5σ, Not 6σ — The Calculation, Origin, and Three Controversies of the 1.5σ Drift Assumption

A week before a customer PPAP review, an automotive parts company was stumped. The customer's drawing stated: key characteristic Cpk≥1.33, "equivalent to 4σ, approximately 63 ppm." On the company's own capability report, the same Cpk 1.33 was listed as "4σ, 6210 ppm." The customer auditor glanced at it and said, "Your data is 100 times worse than the requirement. How can this part be released?"

The purchasing department panicked, assuming the supplier's capability was insufficient. The quality manager checked three training materials and found that they all had different explanations. It was finally clarified: both sides were correct, but one calculated based on a two-sided, no-drift process, while the other used a one-sided, 1.5σ drift process. A single number, two different standards, almost caused a production project to be resampled.

This article will break down the 3.4 ppm conversion chain: how it is calculated, where the 1.5σ comes from, what the controversies are, and which standard should be used in the factory. For the origin of Six Sigma and Bill Smith's repair account, see the previous article. Here, we will focus on the numbers.

1. Conclusion First: 3.4 ppm is "Nominal 6σ, Actual 4.5σ"

In three sentences:

1. When the process is centered and stable, the tail probability of 6σ is 0.00099 ppm on one side, 0.00197 ppm on both sides, approximately 2 ppb, not 3.4 ppm.
2. Shifting the process center by 1.5 standard deviations leaves only 4.5σ on the upper side, with a tail probability of 3.398 ppm—this is the source of 3.4 ppm.
3. Therefore, "Six Sigma = 3.4 ppm" refers to the nominal 6σ level after accounting for the drift margin; the actual tail probability corresponds to 4.5σ. The "σ level" has been a convention since its inception, not a pure probability.

Manual verification (using Excel):

• =NORM.S.DIST(-4.5,TRUE)*1000000 → 3.398
• =NORM.S.DIST(-6,TRUE)*1000000*2 → 0.00197

For reverse conversion, use the engineering approximation formula: Z = 0.8406 + SQRT(29.37 − 2.221 × LN(DPMO)). Substituting DPMO 6210 gives Z≈4.00 (i.e., Cpk 1.33), and substituting DPMO 3.4 gives Z≈6.00—note that this 6.00 is the σ level, not the 4.5 corresponding to the tail probability.

Comparison table (using one-sided standards, which is the common algorithm for Six Sigma tables):

This table can explain most on-site arguments: the customer says "Cpk 1.33 is 4σ," using the column without drift; you say "Cpk 1.33 is 6210 ppm," using the column with drift. Both are correct, but they don't align.

2. How to Determine the Denominator: A 370-Fold Difference in the Same Process

A PCBA factory has 320 solder joints and 48 surface mount components on a single board, totaling 368 defect opportunities per unit. Monthly production is 50,000 units, and inspection records show 214 defects involving 187 nonconforming units. The same data can be written in three ways:

• Nonconforming rate: 187 ÷ 50,000 = 0.374%, or 3740 ppm
• Defect rate (per unit): 214 ÷ 50,000 = 4280 ppm
• DPMO: 214 ÷ (50,000 × 368) = 11.6, which converts to a σ level of Z≈5.73

The highest and lowest values differ by 370 times, and none of the calculations are wrong. Therefore, any statement of "how many ppm we have" must clarify three points: whether the numerator is defects or nonconforming units (see What's the Difference Between Nonconforming Units and Defects), whether the denominator is units, batches, or opportunities, and how the opportunities are defined (whether solder joints are counted repeatedly, whether appearance items are included).

This leads to a hard rule: cross-product comparisons can only be made using DPMO, not ppm nonconforming rate. Complex boards naturally have more opportunities, and using the nonconforming rate for horizontal comparison is like making complex products compensate for simple ones.

At the product level, there is another layer of standards: 3.4 ppm per station, and 200 serial processes result in a defect rate of about 680 ppm for the entire machine. This is precisely why Six Sigma counts "opportunities" rather than "units"—the same overall target, but a coarser definition of opportunities results in an inflated σ level.

3. The Origin of 1.5σ: It Was Originally a Design Margin, Not a Conversion Factor

In the 1980s, Motorola found two things during benchmarking: one, the best companies in the industry were around 6σ, while they were around 4σ; two, the overall standard deviation calculated using data over several months was always larger than the standard deviation calculated using within-group ranges. The difference came from batch changes, mold changes, shifts, environment, and operators—the process center would gradually shift, which is known as "long-term dynamic mean drift."

Their internal empirical value was: long-term Z is about 1.5 lower than short-term Z. Therefore, when they committed to "reaching 6σ (3.4 ppm) in five years" in 1987, the logic was reversed: to ensure that the long-term capability remains 3.4 ppm after five years, they had to design for a short-term 6σ (Cp=2.0, Cpk=1.5). The technical document on tolerance design from the same period used this margin logic.

The original use was very narrow: to leave a margin for time, for target setting and tolerance allocation.

Later, training materials and software extracted the 1.5σ from "design margin" and turned it into a universal conversion table—such as the Z.Bench and DPMO tables in today's software. As the usage expanded, problems arose: the margin became a definition, and the assumption became a fact.

4. Three Controversies: Why Engineers Must Know

Controversy One: 1.5 is an Average, Not an Upper or Lower Limit.

For continuous processes, pharmaceuticals, and enclosed equipment, the measured drift is often between 0.3 and 0.8σ. For multi-variety, small-batch processes with frequent mold changes and manual assembly, a drift of over 2σ is common. Fixing the conversion at 1.5σ is overly pessimistic for the former (turning a real 0.5 ppm into a 3.4 ppm level) and overly optimistic for the latter (the real level might only be slightly above 3σ). Therefore, using 1.5σ for external commitments is risk-asymmetric.

Controversy Two: It Combines "Capability" and "Stability" into One Number.

Cpk is a snapshot of the current state, while drift is a function of time. A process with a Cpk of 1.67 but a control chart showing seven consecutive points rising will definitely degrade over the long term. Conversely, a process that is controlled, shows no out-of-control signals, and has a stable center over 12 months, should not be penalized by a 1.5σ deduction. Voelkel's analysis in ASQ's Quality Progress in 2004 directly questioned whether a drift of 1.5σ is "extremely unlikely." The counter-argument is: without this setting, the target is unachievable. Both sides actually acknowledge that this is an industry convention, not a statistical theorem.

Controversy Three: Different Software Standards Can Yield Two Different Answers for the Same Data.

In Minitab's capability analysis, Z.Within uses the within-group standard deviation, and Z.Overall uses the overall standard deviation, each calculated independently without applying the 1.5σ formula; whereas Excel templates and most training tables directly apply the formula. Using Minitab's Ppk to compare with the 1.5σ table will systematically misalign by one level—this is the technical root of many "report and customer mismatch" issues.

For external communication, don't report the σ level; instead, report three sentences: the first specifies the value and standard—"key characteristic short-term Cpk 1.75, sample spanning 3 shifts"; the second provides an active conversion—"according to your long-term standard, the same data Ppk 1.42, measured drift 0.9σ"; the third suggests a clause change—"we recommend the commitment be written as Ppk≥1.33 and no out-of-control signals on the control chart, rather than '6σ'." Replace unverifiable slogans with recalculable criteria; this is much more useful than arguing whether it should be 3.4 or 0.002.

5. How to Set Standards in the Factory: A Decision Table

Standards should be determined by their purpose, not by habit:

More important than "choosing a standard" is measuring actual drift, not using assumptions. The process is simple:

1. Take 25 groups, each with 5 units (125 points), and calculate the within-group σ using R̄/d₂ (d₂=2.326 when n=5).
2. Calculate the overall standard deviation s using all 125 points.
3. ΔZ = (USL−X̄)/σ_within − (USL−X̄)/s, i.e., Z.Within − Z.Overall.

Criterion: ΔZ≈1.5 indicates that Motorola's convention fits your process; ΔZ≈0.5 suggests that your short-term reports are conservative, and you can communicate at a higher level; ΔZ≥2.0 indicates that all short-term capability reports are inflated, and you should first stratify by batch/mold/shift before discussing the σ level.

A counterexample comes from an injection molding factory: three months of data showed a within-group σ of 0.10 mm, USL−X̄ of 0.50 mm, and a Cpk of 1.67, with the report stating "5σ, 233 ppm." However, the overall s across batches was 0.19 mm, and Z.Overall was only 2.63, with an actual one-sided defect rate of about 4250 ppm, 18 times worse than the report. The root cause of the customer's on-site complaints and long-term report mismatches is that the 1.5σ was treated as a fact, while their actual drift was 2.4σ. For the same batch of data, see The Difference Between Cpk and Ppk and SPC Comprehensive Guide.

6. Self-Checklist

• When the customer document specifies "6σ/3.4 ppm," have I asked whether it is short-term or long-term, one-sided or two-sided?
• Does the footer of our capability report clearly state the formula, standard, and whether the 1.5σ is applied?
• Have we measured the actual drift in our factory using Z.Within − Z.Overall, rather than defaulting to 1.5σ?
• Are we using σ levels for monthly department rankings? (If so, switch to DPMO rankings + control chart signals)
• Before calculating capability, is GR&R ≤10% (or ≤30% and ndc≥5), and is the process under control?

In summary: 3.4 ppm can be a target, but not an actual measurement—first calculate your actual 1.5σ, then decide how to write this report for the customer.

3.4 ppm is the nominal value of 6σ with a 1.5σ drift, not the actual measured probability of 6σ.

Knowledge code: 6.3.2

Version: v20261029

Author: QTank QTank is dedicated to providing systematic professional knowledge, methodologies, and practical tools for quality management practitioners, helping companies continuously improve their quality capabilities.