Operating characteristics of frequently used similarity rules Florian Klinglmueller* Ack: Andreas Brandt, Thomas Lang, Ina Rondak *Austrian Medicines & Medical Devices Agency
The contents of this presentation are my personal opinion. My remarks do not necessarily reflect the official view of AGES.
Introduction Similarity assessment of quality attributes Comparison of an originator product to a biosimilar product with respect to „critical quality attributes“ (i.e. physical, chemical, biological, or microbiological properties that ensure product quality) with the aim to conclude similarity on the quality level. One quality attribute on a continuous scale. Comparison between samples from originator and biosimilar. Different rules to decide whether samples from the biosimilar are similar to samples from the originator - based on sample data Explore operating characteristics of commonly used rules under different scenarios of similarity and dissimilarity.
Decision rule How to compute „ similar/not similar “ from data Problem: Similar is not equal! ⇒ Specify what is similar enough Average similarity • E.g.: Biosim on average within ± 1.5 standard deviations of reference mean • E.g.: Ratio of CQA on average 80%-125% Population based similarity • Reference samples define margin of what is safe (e.g. min-max, TI) • Future/observed biosim batches fall into range Translate into a criterion
Decision rule How to compute „ similar/not similar “ from data Problem: Similar is not equal! ⇒ Specify what is similar enough Average similarity • E.g.: Biosim on average within ± 1.5 standard deviations of reference mean • E.g.: Ratio of CQA on average 80%-125% Population based similarity • Reference samples define margin of what is safe (e.g. min-max, TI) • Future/observed biosim batches fall into range Translate into a criterion
Decision rule How to compute „ similar/not similar “ from data Problem: Similar is not equal! ⇒ Specify what is similar enough Average similarity • E.g.: Biosim on average within ± 1.5 standard deviations of reference mean • E.g.: Ratio of CQA on average 80%-125% Population based similarity • Reference samples define margin of what is safe (e.g. min-max, TI) • Future/observed biosim batches fall into range Translate into a criterion
Statistical Intervals Probabilistic interpretation of different interval types Min-Max: gives a the range of the observed values. Purely descriptive i.e. permits little inference about future samples of the process, except that true range is wider X-SD: estimates the variation in the sample around the sample mean. Purely descriptive, many statistical intervals are constructed by choosing x such that probabilistic statements hold • E.g. 95% CI: x=1.96; 95% PI (n=10): x=2.16, 95/95 TI (n=10): x=3.38 Confidence Interval: estimates a range that should cover an unknown parameter (e.g. mean) of the distribution assumed to generate the data Prediction Interval: estimates a range that should cover the value of the next sample from distribution assumed to generate the data β -content Tolerance Interval: estimates a range that should cover a certain proportion β of future samples from the distribution assumed to generate the data
Statistical Intervals Probabilistic interpretation of different interval types Min-Max: gives a the range of the observed values. Purely descriptive i.e. permits little inference about future samples of the process, except that true range is wider X-SD: estimates the variation in the sample around the sample mean. Purely descriptive, many statistical intervals are constructed by choosing x such that probabilistic statements hold Frequentist confidence: in repeat experimentation range estimate • E.g. 95% CI: x=1.96; 95% PI (n=10): x=2.16, 95/95 TI (n=10): x=3.38 computed in this way will cover the quantity (parameter, next sample, Confidence Interval: estimates a range that should cover an unknown parameter all future samples) a certain proportion of times (e.g. 95%) (e.g. mean) of the distribution assumed to generate the data Prediction Interval: estimates a range that should cover the value of the next sample from distribution assumed to generate the data β -content Tolerance Interval: estimates a range that should cover a certain proportion β of future samples from the distribution assumed to generate the data
Rules for concluding biosimilarity A selection of frequently used decision rules Min-Max: All samples of the biosim are between min-max of the originator X-Sigma: All samples from the biosim are within x-standard deviations of the originators mean (75%/90%) Tolerance interval: All samples from the biosim are within a P/Q Tolerance interval of the originator TI Specs: The P/Q tolerance interval of the biosim is within „specifications“ (e.g. Min-Max) of the originator FDA Rule: The 90% confidence interval for mean difference between originator and biosim is within a similarity margin of 1.5 standard deviations of originator
Simulation scenario 1 Differences in mean, equal variance Originator and Biosim samples follow standard normal distribution Equal variance Distance between distributions expressed as multiples of the (common) standard deviation Settings considered: • M2-M1= {0, 0.5, 0.8, 1, 1.5} * SD
Simulation scenario 1 Differences in mean, equal variance Originator and Biosim samples follow standard normal distribution Equal variance Distance between distributions expressed as multiples of the (common) standard deviation Settings considered: • M2-M1= {0, 0.5, 0.8, 1, 1.5} * SD
Simulation scenario 1 Differences in mean, equal variance Originator and Biosim samples follow standard normal distribution Equal variance Distance between distributions expressed as multiples of the (common) standard deviation Settings considered: • M2-M1= {0, 0.5, 0.8, 1, 1.5} * SD
Simulation scenario 1 Differences in mean, equal variance Originator and Biosim samples follow standard normal distribution Equal variance Distance between distributions expressed as multiples of the (common) standard deviation Settings considered: • M2-M1= {0, 0.5, 0.8, 1, 1.5} * SD
Simulation results: Equal variances Most simple scenario Equal means (biosimilar) shown with dashed lines, unequal means solid lines. Probability to conclude similarity decreases with increasing dissimilarity (m=10,n=10)
Simulation results: Equal variances Most simple scenario Equal means (biosimilar) shown with dashed lines, unequal means solid lines. Probability to conclude similarity decreases with increasing dissimilarity (m=10,n=10) If we increase the sample size (m=20,n=20) several things happen
Simulation results: Equal variances Most simple scenario Equal means (biosimilar) shown with dashed lines, unequal means solid lines. Probability to conclude similarity decreases with increasing dissimilarity (m=10,n=10) If we increase the sample size (m=20,n=20) several things happen Similarity conclusions increase, with increasing originator samples (except TI)
Simulation results: Equal variances Most simple scenario Equal means (biosimilar) shown with dashed lines, unequal means solid lines. Probability to conclude similarity decreases with increasing dissimilarity (m=10,n=10) If we increase the sample size (m=20,n=20) several things happen Similarity conclusions increase, with increasing originator samples (except TI) With increasing Biosim samples similarity conclusions increase for for interval based criteria
Simulation scenario 2 Difference in means, unequal variance Originator and biosim distribution may differ in mean and in variance Standard deviation of originator is sdratio times larger than biosim Values for sdratio: .25 - 2 sdratio=1 corresponds to Scenario 1 Both cases, assuming equal and unequal means, were investigated
Simulation scenario 2 Difference in means, unequal variance Originator and biosim distribution may differ in mean and in variance Standard deviation of originator is sdratio times larger than biosim Values for sdratio: .25 - 2 sdratio=1 corresponds to Scenario 1 Both cases, assuming equal and unequal means, were investigated
Simulation scenario 2 Difference in means, unequal variance Originator and biosim distribution may differ in mean and in variance Standard deviation of originator is sdratio times larger than biosim Values for sdratio: .25 - 2 sdratio=1 corresponds to Scenario 1 Both cases, assuming equal and unequal means, were investigated
Simulation scenario 2 Difference in means, unequal variance Originator and biosim distribution may differ in mean and in variance Standard deviation of originator is sdratio times larger than biosim Values for sdratio: .25 - 2 sdratio=1 corresponds to Scenario 1 Both cases, assuming equal and unequal means, were investigated
Simulation scenario 2 Difference in means, unequal variance Originator and biosim distribution may differ in mean and in variance Standard deviation of originator is sdratio times larger than biosim Values for sdratio: .25 - 2 sdratio=1 corresponds to Scenario 1 Both cases, assuming equal and unequal means, were investigated
Recommend
More recommend
