Correlations with critical-to-quality product deviations
Using critical-to-quality defects at the commercial scale as the dependent variable, 56 other variables were compared based
on likelihood ratios for each of the 34 commercial products. Of these 56 variables, 11 were found to have statistically significant
different likelihoods depending upon whether a critical-to-quality defect was experienced for that product. These 11 variables
of significance are shown in Figure 4 and grouped based on confidence level of statistical significance (the top row contains variables for which the difference
in likelihood ratio demonstrated greatest confidence) and grouped based on whether the correlation was negatively or positively
correlated with a critical-to-quality defect at the commercial scale. If the variable is in the negatively correlated section
of Figure 4, then presence of that variable is more likely to occur for a product without a reported critical defect.
Table I: Variables that were found to be statistically different in likelihood of correlating with a critical-to-quality defect.
For example, a product manufactured at commercial scale in a site having an above-median fraction of technical personnel working
in QA was statistically less likely to have experienced a critical-to-quality defect at the commercial scale. On the other
hand, a positively correlated variable is more likely to be present for a product that has experienced a critical defect.
For example, a product manufactured in Asia is more likely to have experienced such a defect.
Consistent with prior preliminary discussion regarding variables correlated with product defects, many of the statistically
significant variables presented in Figure 4 fall into the categories of history of quality issues in the product lifecycle, geographic region of manufacture, and manufacture
in a contract manufacturing site (2). Of note is the dependence upon fraction of technical personnel employed in QA and employed
in manufacturing. More generally, this dependence is related to the discussion of the distribution of technical personnel
across a site as described above, where the lack of consistency and, perhaps, understanding of an optimal distribution was
highlighted. The data presented in Figure 4 demonstrates that above median level of the fraction of technical personnel in QA and below median level of the fraction
of technical personnel in manufacturing are both correlated with lower likelihood of a critical-to-quality defect at the commercial
It is possible that greater emphasis on quality is demonstrated by fraction of personnel in QA compared with an emphasis on
execution demonstrated by fraction of personnel in manufacturing. Further investigation in this area and clarity regarding
the expectations of technical personnel roles could provide greater insight into this observation.
Although there were 11 variables significantly correlated with likelihood of a critical-to-quality defect at the commercial
scale, there were also 45 variables found not to be so. Several of these variables are notable as not having been found to
correlate with critical-to-quality defects at the commercial scale, including:
Type of product (e.g., protein, antibody).
Mode of production (e.g., mammalian, microbial).
Level of production complexity perceived by the site: The first three variables are related to perceived manufacturing complexity and logic could suggest that a dependence is
expected, but has not been found in the current data.
Use of QbD: QbD can take on many forms and has a complex impact on measured quality performance metrics. Generally, this lack of correlation
is notable as quality by design is expected to increase the likelihood of product with consistent quality over time (3).
Number of product batches manufactured in the past five years: More batches would be expected to increase probability of a single defect occurring but is logically counteracted by familiarity
with the product and the possibility of increased learning from completed batches (4).