Why Use a Similarity Factor?
Two dissolution profiles can look close overall while differing at particular times, or look different simply because measurements vary among dosage units. A numerical summary can help describe their average separation across time, but it is useful only when the profiles are sampled comparably and their variability is not too high for the method. The \(f_2\) similarity factor is a conventional summary for this purpose.
The factor compresses the squared differences between the test and reference mean dissolution percentages at matched time points into one number. It is not a statistical test, does not quantify uncertainty, and does not replace checking the profile and the conditions for using the calculation.
An \(f_2\) value of 100 occurs when the two mean profiles are identical at every included time point. As their average squared separation increases, \(f_2\) decreases. Under the conventional criterion, values from 50 to 100 indicate similar profiles, provided the calculation’s conditions are met. This is a rule for interpreting the summary, not a claim that the products are bioequivalent or that every part of their profiles is identical.
Conditions for a Defensible Calculation
The usual \(f_2\) procedure compares mean dissolution profiles from individual dosage units. Conventional recommendations commonly call for at least 12 units of each product, at least three nonzero time points, identical sampling times for the two products, and no more than one included time point after both products have exceeded 85% dissolved. The zero-time value is not used: both products begin at zero, so including it adds no information about profile separation. Applicable product-specific or regulatory guidance may set different or additional requirements.
Variability is checked at each time point from the unit-level measurements, rather than inferred from the means alone. A common conventional criterion is a coefficient of variation no greater than 20% at earlier time points (for example, up to 10 minutes) and no greater than 10% at later time points. For a set of unit measurements, the coefficient of variation is \(100s/\bar{x}\), expressed as a percentage, where \(\bar{x}\) is the sample mean and \(s\) the sample standard deviation. The time boundary and applicable limits should follow the relevant guidance. If variability exceeds the criterion, the numerical factor can still be computed, but its usual similarity interpretation is not justified by that criterion.
Time points should represent the same elapsed times for both products. Pairing a reference result at 15 minutes with a test result at 20 minutes does not meet that requirement. Nor should many late measurements from a nearly complete-release plateau dominate the summary: the conventional restriction on points after 85% helps prevent that. Before calculating \(f_2\), establish which points are eligible and check that the same points are used for both products.
Worked Example: Calculating f2 From Matched Profiles
Suppose 12 units of each hypothetical product were tested at the same four nonzero times. The unit-level variability meets the applicable conventional limits, and the final time is the only included time after both mean profiles exceed 85%. The mean percentages dissolved are:
| Time (minutes) | Reference mean (%) | Test mean (%) | Difference \(R_t-T_t\) (percentage points) | Squared difference |
|---|---|---|---|---|
| 5 | 20 | 18 | 2 | 4 |
| 10 | 45 | 48 | -3 | 9 |
| 15 | 70 | 68 | 2 | 4 |
| 30 | 88 | 91 | -3 | 9 |
There are \(n=4\) matched points. The mean squared difference is \((4+9+4+9)/4=6.5\) squared percentage points. Substituting this value into the definition gives
The calculated value is above 50. Because the example also satisfies the stated sampling and variability conditions, the conventional interpretation is that the two mean dissolution profiles are similar by the \(f_2\) criterion. This conclusion is limited to this profile-comparison rule; it does not establish in vivo bioequivalence.
Why the Threshold Has Its Particular Meaning
The threshold can be related directly to the mean squared difference. Let \(M\) denote the mean of the squared differences across the \(n\) included points. By definition, \(M\geq 0\), and
Because the logarithm is increasing, \(f_2\geq 50\) precisely when its argument is at least \(10\). Solving this inequality gives
Thus, at the conventional cutoff, the mean squared difference across the included points is no greater than 99 squared percentage points. This algebra explains the cutoff; it does not make \(f_2\) a test of a population hypothesis. In particular, it does not account for the uncertainty in estimated mean profiles.
Worked Example: Checking Variability Before Interpreting f2
In a second hypothetical study, 12 units of each product are measured at 5, 10, 15, and 30 minutes. The table gives the mean profiles and the coefficients of variation calculated from the unit-level results. Assume the conventional boundary is 10 minutes: 5 and 10 minutes are earlier points, while 15 and 30 minutes are later points.
| Time (minutes) | Reference mean (%) | Test mean (%) | Reference CV (%) | Test CV (%) |
|---|---|---|---|---|
| 5 | 24 | 22 | 12 | 15 |
| 10 | 51 | 50 | 16 | 18 |
| 15 | 76 | 74 | 8 | 9 |
| 30 | 91 | 90 | 5 | 6 |
The early-point CVs are all no greater than 20%; the later-point CVs are all no greater than 10%. The time points match, and there is only one included point after both means exceed 85% (30 minutes). With 12 units per product and four nonzero points, the example meets the conventional conditions stated here.
The differences are 2, 1, 2, and 1 percentage points, so the mean squared difference is \((4+1+4+1)/4=2.5\). Therefore,
The value exceeds 50, and the checks support interpreting it as similar by the conventional \(f_2\) criterion. The CV check matters: the same means and calculated factor would not, by themselves, demonstrate that the method’s variability condition was met.
What the Number Does—and Does Not—Show
The calculation gives equal weight to each included time point’s squared difference. A large difference at one time point can therefore reduce \(f_2\), even when the other points are close. Conversely, a high factor can conceal a localized difference if the selected time points do not capture it. The factor also uses means, so it does not describe the spread of individual-unit profiles or reveal whether some units behave quite differently from the average.
A simple visual comparison is useful for seeing the shape, direction, and timing of differences, but it does not produce a single conventional summary. The difference factor \(f_1\) is another descriptive measure, based on absolute rather than squared differences; it can complement profile inspection, but it does not remove the need to check sampling and variability conditions. Neither \(f_1\) nor \(f_2\) is a substitute for an uncertainty analysis when the question requires one.
Interpret \(f_2\) as a conditional summary of the sampled mean profiles—not as a probability, a confidence interval, a significance test, or a general guarantee of equivalent product performance. Its result depends on the chosen time points, and changing that selection can change the factor. A value just above 50 should not be treated as strong evidence of identity; it merely meets the conventional threshold under the conditions used.
Worked Example: A High Factor With a Variability Problem
Consider a third hypothetical comparison with 12 units per product at 5, 10, and 20 minutes. The mean percentages are 30, 60, and 90 for the reference, and 28, 62, and 91 for the test. The CVs at 5 minutes are 12% and 13%; at 10 minutes, 18% and 19%; and at 20 minutes, 14% and 12%, respectively. Assume 5 and 10 minutes are the earlier points and 20 minutes is a later point. There is one point after both means exceed 85%, and the times match, but both CVs at 20 minutes exceed the conventional 10% later-point limit.
The differences are 2, -2, and -1 percentage points. The mean squared difference is \((4+4+1)/3=3\), giving
The number is greater than 50, but the later-point variability condition is not met. It would be incorrect to report that the profiles have been shown similar by a valid conventional \(f_2\) comparison. The appropriate conclusion is that the calculated summary is 84.95, but its standard similarity interpretation is not supported under the stated variability criterion. The variability issue needs to be investigated using the relevant guidance and study context.
Common Mistakes and Practical Checks
- Confusing time points with units. In the formula, \(n\) counts matched time points. The number of dosage units informs the mean and the variability check; it is not substituted for \(n\).
- Pairing different sampling times. A reference value at 15 minutes and a test value at 20 minutes are not a matched pair. Use the same elapsed times, rather than silently treating unequal times as equivalent.
- Using only mean profiles to check variability. CV requires unit-level measurements at each time. Similar means do not guarantee that the variability condition is met.
- Including many late plateau points. Repeated points after both profiles exceed 85% can overweight the near-complete-release region. Follow the applicable restriction when choosing included points.
- Calling \(f_2\geq 50\) proof of bioequivalence. The cutoff is a conventional dissolution-profile similarity criterion, not a clinical bioequivalence test.
A useful sequence is to verify the unit count and sampling schedule, identify the eligible time points, check unit-level CVs, calculate each paired difference, and then compute and interpret \(f_2\). If any key condition fails, report that limitation rather than presenting the threshold as an unconditional conclusion.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Why is the number of time points, rather than the number of dosage units, used as \(n\) in the \(f_2\) formula?
- Two profiles have \(f_2=54\), but the test product’s CV at a later time point is 13%. Under the conventional 10% later-point limit, what conclusion is justified?
- For three matched points, the squared differences are 1, 4, and 4. Calculate \(f_2\) and interpret it if all stated conditions are met.
- Why should a reference measurement at 15 minutes not be paired with a test measurement at 20 minutes?
Solutions. (1) \(n\) is the number of paired time-point differences being averaged; dosage units provide the observations used to estimate each mean and CV. (2) The numerical value is above 50, but the later-point variability condition fails, so the conventional similarity interpretation is not justified. (3) \(M=(1+4+4)/3=3\); thus \(f_2=50\log_{10}(100/\sqrt{4})=50\log_{10}(50)\approx84.95\), which indicates similarity by the conventional criterion when its conditions are met. (4) The values represent different elapsed times and are not a matched pair, so their difference does not compare the products at the same point in the dissolution process.