When a Single Similarity Number Is Not Enough
The f2 similarity factor compresses the differences between two mean dissolution profiles into one number. This is useful when the profiles are measured on a suitable common schedule and their variability is acceptable for the method. But the number can create false reassurance if the measurements are highly variable, the sampling schedule misses important changes, or the profiles diverge at particular times and then cross.
The right response is not automatically to replace f2 with another summary. First identify what could be misleading, then choose a comparison that answers the study question and is supported by the data. A numerical f2 value is descriptive; by itself, it does not quantify uncertainty or establish that two formulations are equivalent.
A value of \(f_2=100\) occurs when the two mean profiles agree at every included time. Smaller values indicate greater average differences. A commonly used convention treats \(f_2\geq 50\) as consistent with profile similarity, but that convention is meaningful only when the method and data meet the relevant conditions. Regulatory guidance and the product context determine which conditions apply.
What the Formula Can and Cannot Tell You
The formula averages squared time-point differences. Squaring prevents positive and negative differences from canceling algebraically, but averaging still hides when the differences occur and whether they form a systematic shape. It also gives each included time point equal weight, regardless of the time interval since the previous observation or the clinical importance of release during that interval.
One useful justification follows directly from the formula. Let \(M\) be the mean squared difference, \(\frac{1}{n}\sum_{t=1}^{n}(R_t-T_t)^2\). For the conventional threshold \(f_2\geq 50\), $$ 50\log_{10}\left(\frac{100}{\sqrt{1+M}}\right)\geq 50 \quad\Longleftrightarrow\quad \frac{100}{\sqrt{1+M}}\geq 10 \quad\Longleftrightarrow\quad M\leq 99. $$ So the threshold corresponds to a root-mean-square difference no greater than \(\sqrt{99}\), or about 9.95 percentage points, across the selected time points. This is an average over the chosen schedule, not a guarantee that every time-point difference is small.
In conventional applications, the profiles should be measured under the same dissolution conditions and compared at the same nonzero time points. Common regulatory conventions also call for an adequate number of dosage units, typically at least 12 per formulation, and a suitable set of sampling times. Often-cited variability checks are that the coefficient of variation is no more than 20% at early time points and no more than 10% at later time points. The applicable guidance should be checked for the product and jurisdiction. When these variability checks are not met, a standard f2 interpretation may not be reliable.
Worked Example: A Passing f2 With High Unit Variability
In a hypothetical study, 12 units of each formulation are tested at four common times. The mean reference profile is 20%, 45%, 70%, and 90% dissolved; the mean test profile is 22%, 47%, 72%, and 90%. Suppose the corresponding unit-level standard deviations are 12, 18, 18, and 9 percentage points for the reference, and 13, 17, 20, and 12.6 percentage points for the test.
Calculate f2. The mean differences are \(-2,-2,-2,0\) percentage points. Therefore, $$ f_2=50\log_{10}\left(\frac{100}{\sqrt{1+\frac{(-2)^2+(-2)^2+(-2)^2+0^2}{4}}}\right) =50\log_{10}\left(\frac{100}{\sqrt{4}}\right) =50\log_{10}(50) \approx 84.95. $$
The numerical value is above 50, but that does not resolve the variability concern. For example, at the first time the reference coefficient of variation is \(12/20\times100=60\%\); at the final time the test coefficient of variation is \(12.6/90\times100=14\%\). Both exceed the commonly cited checks for their respective time regions. The f2 calculation uses the mean differences and does not incorporate these standard deviations. The defensible conclusion is that the calculated f2 is 84.95, but the conventional similarity interpretation is not supported without addressing the high variability under the applicable guidance.
Sampling Times Can Change the Answer
A sampling schedule is part of the comparison, not a neutral detail. If observations are concentrated after both products have nearly completed dissolution, substantial differences during the rising portion of the profiles can go unseen. Conversely, a dense schedule in a region where profiles differ briefly can lower f2 even if they are close at other times. Choosing times after looking at the observed differences can also make the comparison difficult to defend; sampling should be designed to characterize the expected profile and relevant transitions.
Worked Example: A Sparse Schedule Misses a Divergence
Suppose hypothetical mean profiles, in percent dissolved, are as follows. If only 1, 4, and 8 hours are sampled, the reference and test means are both 20%, 70%, and 95%.
| Time (hours) | Reference (%) | Test (%) | Test minus reference (percentage points) |
|---|---|---|---|
| 1 | 20 | 20 | 0 |
| 2 | 45 | 30 | -15 |
| 4 | 70 | 70 | 0 |
| 6 | 88 | 96 | 8 |
| 8 | 95 | 95 | 0 |
Using only 1, 4, and 8 hours, every included difference is zero, so \(f_2=100\). That result says only that the means agree at those sampled times.
Adding the 2- and 6-hour observations gives five differences: \(0,-15,0,8,0\). The mean squared difference is \((0^2+(-15)^2+0^2+8^2+0^2)/5=289/5=57.8\) percentage-points squared. Thus, $$ f_2=50\log_{10}\left(\frac{100}{\sqrt{1+57.8}}\right) =50\log_{10}\left(\frac{100}{\sqrt{58.8}}\right) \approx 55.77. $$
The denser schedule reveals a substantial transient divergence that the sparse schedule missed. The two values are not contradictory: they summarize different selected time points. The diagnostic lesson is to inspect whether the planned times can capture the expected rising phase, any lag or transition, and the approach to the plateau.
Crossing Profiles and Local Differences
When one profile is above the other at some times and below it at others, the curves cross. Because f2 squares differences, opposite signs do not cancel. Yet its single average still loses the direction, timing, and ordering of those differences. A crossing can matter when an early release difference has a different interpretation from a late one, or when a specific part of the profile is important to the study question.
Worked Example: Similarity Threshold With Alternating Differences
Consider hypothetical means at five common times. The reference profile is 20%, 40%, 60%, 80%, and 95%; the test profile is 30%, 30%, 70%, 70%, and 95%. The test-minus-reference differences are \(10,-10,10,-10,0\) percentage points, so the profiles cross repeatedly.
Their mean squared difference is \((100+100+100+100+0)/5=80\) percentage-points squared. Therefore, $$ f_2=50\log_{10}\left(\frac{100}{\sqrt{1+80}}\right) =50\log_{10}\left(\frac{100}{9}\right) \approx 52.29. $$
This passes the commonly used numerical threshold, but the largest absolute difference is 10 percentage points, and the direction of the difference changes over time. Whether that pattern is acceptable cannot be decided from f2 alone. A defensible interpretation would show the time-specific differences and relate them to prespecified, scientifically justified criteria; there is no universal pointwise acceptance margin that can be inferred from the f2 threshold.
Diagnostics and Alternative Comparisons
Before relying on f2, examine the data at both the profile and unit levels. Plot the mean profiles together, then plot the time-specific test-minus-reference differences. Where possible, show individual-unit trajectories or distributions at each time. Look for large or changing coefficients of variation, outlying units, irregular shapes, apparent crossings, and time regions with too few observations to characterize the release pattern. A smooth-looking mean curve does not demonstrate that individual units behave consistently.
When a diagnostic identifies a limitation, select an alternative that matches the question rather than searching for a statistic that gives a preferred conclusion. Main options have different strengths and limitations:
| Approach | What it emphasizes | Main limitation |
|---|---|---|
| Time-specific differences | Direction and size of the test-reference difference at each time | Many separate comparisons require a prespecified interpretation and careful treatment of uncertainty. |
| Repeated-measures or other unit-level model | Differences over time while accounting for within-unit observations and variability | Requires a justified model, adequate data, and diagnostics; a model does not repair a poor sampling design. |
| Integrated summaries, such as area under the dissolution curve | Overall cumulative release across a defined time interval | Different local profiles can have similar areas, so timing differences may be hidden. |
| Model-based profile comparison | Differences in interpretable curve features or predicted profiles | Depends on model adequacy and should not be used to extrapolate beyond the measured data without justification. |
For inferential comparisons, uncertainty should be quantified and the decision criteria specified before the results are interpreted. If the scientific question is equivalence within a margin, that margin must be justified independently; a nonsignificant difference test is not evidence of equivalence. Integrated or model-based summaries can be useful, but neither should replace a clear statement of what aspect of the profile is being compared.
No summary can compensate for sampling times that miss the feature of interest or for inadequate replication when unit variability is substantial. If the design is the source of the problem, the most defensible alternative may be to collect better data rather than apply a more elaborate analysis to the same limited observations.
Common Pitfall
Wrong reasoning: “The f2 value is above 50, so the profiles are similar at every time and the formulations are equivalent.”
Correction: The threshold describes an average across selected time-point means under suitable conditions. It does not show that every difference is small, account for high unit variability, establish equivalence, or protect against unsampled divergences. Check the schedule, variability, and shape of the profiles, then use an analysis suited to the specific question.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions.
- Why can a high f2 value fail to address high variability among dosage units?
- Two profiles agree at every sampled time. What does \(f_2=100\) establish, and what does it not establish?
- Why does a crossing pattern require more than reporting one f2 value?
- Which comparison would best emphasize overall release across a specified interval, and what feature could it conceal?
Solutions. (1) f2 is calculated from differences between time-point means and does not directly use unit-level standard deviations. High variability can make those means an inadequate basis for a conventional similarity conclusion. (2) It establishes agreement of the included means at the included times; it does not establish agreement between those times, low unit variability, or equivalence. (3) The average hides the timing and direction of local differences, which may matter to the scientific question. Plot and report time-specific differences, and apply prespecified criteria where appropriate. (4) An integrated summary such as area under the dissolution curve emphasizes cumulative release over the stated interval, but can conceal local timing differences or different curve shapes.