How do you avoid confirmation bias in a root cause investigation?

This method covers the bias safeguards in a quality or performance complaint about a chemical product. How to run the whole root cause analysis, and incident investigations with injuries or releases, follow their own procedures and are covered elsewhere.

Why do root cause investigations lock onto the first suspect?

Root cause investigations lock onto the first suspect because early information carries more weight than later information, and investigators then read new data in a way that fits the first idea. Nickerson defines confirmation bias as "the seeking or interpreting of evidence in ways that are partial to existing beliefs, expectations, or a hypothesis in hand" (Nickerson, 1998). He calls the early-weighting part the primacy effect.

In a chemical product complaint, the first suspect usually arrives with the complaint itself. The customer says the new lot is bad, or the sales engineer remembers a raw material change, and the investigation starts as a search for proof. Heuer describes the same pattern in analysts: they pick the answer they suspect is most likely and look at the information "from the point of view of whether or not it supports this answer" (Heuer, 1999).

The trap is that most evidence fits several explanations. A coating that lost adhesion fits a bad lot, a contaminated substrate, a short cure and a cold workshop at the same time (illustrative). Heuer's warning is that analysts "fail to recognize that most of this evidence is also consistent with other explanations" when they work one hypothesis at a time (Heuer, 1999). A procedure that lists rivals before testing is the practical answer, because it gives the other explanations a written place before the first result arrives.

How do you avoid confirmation bias in a root cause investigation, step by step?

You avoid confirmation bias in a root cause investigation with four steps done in order, each one written down before the next begins. The steps follow Heuer's competing hypotheses and Platt's strong inference, adapted to a product complaint (Heuer, 1999; Platt, 1964). They sit inside a full root cause analysis process such as the one IEC 62740:2015 describes, not in place of it (IEC 62740:2015).

1. List rival hypotheses. Write at least one candidate in each cause area: the product lot, storage and handling, the customer's application conditions, the substrate or system, and contamination. Heuer's step 1 asks for a group of analysts with different perspectives for this, because one person rarely thinks of every possibility (Heuer, 1999).

2. Fix what would disprove each. For every hypothesis, write the observation or test result that would rule it out, and the result you expect if it is true.

3. Test. Run the checks that separate the hypotheses, starting with those whose outcomes exclude one or more candidates, as Platt describes a crucial experiment (Platt, 1964). Keep the number of repeats fixed in advance.

4. Record excluded causes with scope. For each hypothesis ruled out, write which lots, dates and conditions the exclusion covers, and which evidence did it.

Hypotheses that are neither confirmed nor ruled out stay open. Heuer is explicit: "Unproven hypotheses should be kept alive until they can be disproved" (Heuer, 1999).

How do you fix root cause hypotheses and disconfirming conditions before collecting evidence?

You fix root cause hypotheses and disconfirming conditions before collecting evidence by writing, for each candidate cause, one prediction and one result that would rule it out, dated and agreed before the first test result is read. The point is that the rule for rejecting a hypothesis cannot move once the data starts to arrive.

A disconfirming condition names a test, a sample and a threshold. For a "bad lot" hypothesis behind a viscosity complaint, it could read: "If the retained sample of lot 2409-12 meets the certificate of analysis viscosity range when tested by the same method, lot composition is excluded as the cause of the viscosity complaint" (illustrative). For a contamination hypothesis it could name the water content of the customer's in-use sample (illustrative).

Regulated laboratories already work this way for retests. FDA guidance for pharmaceutical production calls repeated testing until a passing result appears "testing into compliance" and says it "is unscientific and objectionable under CGMP" (FDA, 2022). The same guidance says the number of retests "should not be adjusted depending on the results obtained" (FDA, 2022). The guidance is written for pharmaceutical CGMP laboratories; the habit of fixing the stopping rule first transfers to any chemical product investigation.

Store the hypothesis list with its date. If a new hypothesis appears later, add it with its own date instead of rewriting the original list.

How do you record excluded causes and the scope of each exclusion?

You record excluded causes by writing, for every ruled-out hypothesis, the evidence that excluded it and the exact scope of that exclusion: the lot range, the time window and the condition tested. An exclusion without scope reads as "not the product", when the evidence may only show "not lot 2409-12, on viscosity, at 40 °C" (illustrative).

An exclusions log needs five fields per entry:

• Hypothesis: the candidate cause, worded as it was fixed before testing.

• Evidence: the test, sample and result, with the document it came from.

• Lot range: which lots or batches the evidence covers, and which it does not.

• Time window: the production, delivery or use dates the evidence covers.

• Condition tested: the property, method and conditions, such as temperature or concentration.

The scope matters because the next complaint may fall outside it. FDA guidance for pharmaceutical laboratories says that even after the root cause is found, a failure investigation that extends to "other batches or products that may have been associated with the specific failure" must be completed (FDA, 2022). An exclusion recorded with its lot range and time window shows which other lots were covered and which still need a check.

Keep hypotheses that were never tested in the log as open, not excluded. Heuer separates disproved hypotheses, which have "positive evidence that it is wrong," from unproven ones that simply lack support (Heuer, 1999).

Why do root cause methods disagree on forming hypotheses before collecting facts?

Root cause methods disagree on forming hypotheses before collecting facts because they guard against different failures: fact-first methods guard against a premature favorite, and hypothesis-first methods guard against collecting everything without knowing what would decide the question. Both sides aim at the same bias.

TapRooT states that its method avoids confirmation bias "by first requiring the collection of information to thoroughly understand what happened before starting to find out why it happened" (TapRooT, 2019). Its article on unbiased investigations warns that investigators "seek out, interpret, and remember information that aligns with their initial hypotheses" (TapRooT, 2023).

Heuer and Platt take the other order. Heuer's method starts by identifying the possible hypotheses and then tests them against the evidence (Heuer, 1999). Platt's strong inference starts by "Devising alternative hypotheses" and then designs experiments to exclude them (Platt, 1964). An engineering.com article on the Red X approach describes data "collected by testing a variety of hypothesis" (engineering.com, 2023).

The two orders meet at one safeguard. Preserve and record the facts first, as FDA guidance does when it asks for the initial assessment before test preparations are discarded (FDA, 2022). Then list several hypotheses, and fix the disconfirming condition for each before the tests that decide them. A single early hypothesis is the risk in both methods, not the timing itself.

How do you check a root cause investigation for confirmation bias?

You check a root cause investigation for confirmation bias by testing its records against four questions: were rival hypotheses listed, was a disconfirming condition fixed for each before testing, was the number of tests fixed in advance, and does every exclusion state its scope. The table applies these to a chemical product complaint.

The table below decides whether a root cause investigation tested rival hypotheses or only confirmed the first suspect.

The exclusion scope row is the one most investigations fail: a log that says "product excluded" has to be rewritten to name the lots and properties tested, or the next complaint will reuse an exclusion that never covered it. Run the checks before the conclusion goes to the customer, while tests can still be repeated on retained samples. Lubinpla runs root cause analysis on the same rule, fixing hypotheses and their disconfirming conditions first and recording each excluded cause with the scope of the exclusion.