Wire-Bond Failures: Confounded Shift, Machine, and Lot Effects
A semiconductor assembly line produces a plastic package with a 6-mil gold wire bond. Final test shows 4.2% pull-strength failures against a lower spec of 3.0 gf. Data from the last eight weeks include three shifts (A, B, C) and four bonders (1-4). Shift A has 6.8% failures, B 3.1%, C 2.7%. Bonder 2 has 7.4% failures and ran only on shift A during the period; bonder 4 ran on all shifts with 3.0% failures. A new operator training module was introduced in week 5 for shift A only. The team's first instinct is to conclude shift A is the root cause and retrain all shifts. The process engineer notes that bonder 2 also had a changed capillary supplier in week 3, and the failure rate on bonder 2 rose after week 4. Weekly production volume is 18,000 units, split unevenly across shifts. The belt has unit-level pass/fail data, machine ID, shift, operator ID, capillary lot, and week.
1. Assess whether the shift effect is confounded, and design an analysis to separate shift, machine, and capillary lot.
Read the worked answer
The shift comparison is confounded and should not be interpreted as a shift effect. Bonder 2 ran only on shift A and has the highest failure rate (7.4%), so the shift A rate of 6.8% is a mixture of machine and shift. The capillary supplier change in week 3 and the training module in week 5 add time confounding: the rise on bonder 2 after week 4 could be capillary lot, training, or both. A multi-vari chart of failure rate by week, machine, shift, and capillary lot is the first step. Then fit a logistic regression with machine, shift, capillary lot, week, and operator; check for machine-by-shift interaction. If bonder 2 is completely nested within shift A, no statistical model can separate the two; the team needs data from bonder 2 on another shift or a designed run. Report cell counts and confidence intervals, not just overall percentages. This prevents a Simpson's paradox conclusion.
2. What does the evidence suggest, and what additional data or action is needed before retraining all shifts?
Read the worked answer
The evidence points more strongly to bonder 2 and the capillary change than to shift A as a whole. Bonder 4 ran on all shifts at 3.0%, which shows that shift alone does not produce the high failure rate. Bonder 2's 7.4% and its rise after week 4 coincide with the capillary supplier change. The training module could also have affected shift A, but it was introduced in week 5, after the capillary change. Before retraining all shifts, the team should: trace capillary lots to bonder 2 by week; run bonder 2 with an old and a new capillary lot under controlled conditions; compare shift A bonders 1, 3, and 4 before and after training; and collect data when bonder 2 runs on another shift if production allows. If bonder 2 remains high across shifts, the root cause is machine/capillary; if only shift A operators are high across machines, training is more plausible. Retraining everyone now would be a costly action based on a confounded signal.
3. How would you communicate the risk of the team's initial conclusion?
Read the worked answer
Communicate with a simple stratification: show the overall shift rates, then the machine-within-shift rates. Explain that shift A's high rate is driven by one machine that never runs on other shifts, so the data cannot tell us whether the cause is the shift or the machine. State the two competing hypotheses and the evidence for each. Recommend a short, low-cost confirmation: 2 weeks of capillary lot traceability plus a small run on bonder 2 with controlled capillary lots, and if possible a temporary schedule that places bonder 2 on shift B for a few lots. Do not approve enterprise-wide retraining until the confirmation is complete. Update the FMEA with capillary lot and machine as potential causes, and add a change-control check for supplier changes. This is a risk-communication and decision-quality issue: the team should avoid acting on an unseparated effect even if the sponsor wants a quick fix.