The higher benefit of lecanemab in males compared to females in CLARITY AD is probably due to a real sex effect

Abstract

INTRODUCTION

The phase 3 trial CLARITY AD found lecanemab slowed cognitive decline by 27%. However, subgroup analyses indicated a significant 31% sex difference in the effect and suggested no or limited effectiveness in females. We used simulations constrained by the trial design to determine whether that difference reflects a pre‐existing sex difference in Alzheimer's disease progression or was a random event.

METHODS

Simulations were generated using linear mixed models of cognitive decline fit to data from Alzheimer's Disease Neuroimaging Initiative participants satisfying CLARITY AD inclusion criteria.

RESULTS

The statistically non‐significant 7.9% smaller cognitive decline rate in our cohort's males versus females does not explain CLARITY AD's 31% sex difference in lecanemab's effect. A ≥ 31% difference occurred randomly in only 12 of our 10,000 simulations (0.0012 probability).

DISCUSSION

CLARITY AD's sex difference was probably not random. Lecanemab is likely less effective in females than males, but we cannot conclude the drug is ineffective in females.

Highlights

• Lecanemab is more clinically effective in males than in females.

• Forest plots should only report subgroup‐specific effects in well‐powered subgroups.

• Trial simulations based on real data enable investigation of subgroup drug effects.

• We cannot conclude that lecanemab is clinically ineffective in females.

• A sex difference in lecanemab's efficacy could be linked to its action mechanism.

1. BACKGROUND

Alzheimer's disease (AD) affects males and females at different rates. ¹ Two thirds of patients are female, and lifetime risk in females is twice that of males. ² , ³ , ⁴ Sex differences in genetics, hormones, and other biological mechanisms are hypothesized to contribute to sex differences in the disease progression and, possibly, drug efficacy. ² Recent work has therefore called for clinical trials powered to include primary endpoints disaggregated by sex. ² , ⁵ While no such trial has yet been done in AD, trials of new drugs do often run sex subgroup analyses. ⁶ , ⁷ , ⁸

The purpose of such analyses is to generate hypotheses for future research, not to evaluate efficacy within each subgroup. ⁹ Ideally, statistical interaction tests are used to determine whether one subgroup benefited more from a drug than a corresponding subgroup. ⁹ In some trials, however, subgroup results are only reported visually on a forest plot, where the primary endpoint and confidence interval are presented for each subgroup individually. ⁶ , ⁷ , ¹⁰ This format implies that subgroup‐specific effectiveness was evaluated, despite typical subgroup sizes not being powered to test the primary hypothesis. ⁶ , ⁷ , ⁸ , ⁹ Biostatisticians recommend that primary hypothesis tests not be run and reported separately for low‐powered subgroups. ⁹ , ¹⁰ , ¹¹

These complications make forest plots hard to interpret clinically without strong statistical expertise, especially if there are striking differences between corresponding subgroups. That exact scenario may have contributed to important misinterpretations of the sex subgroup results in CLARITY AD, the phase 3 trial of the amyloid‐targeting drug lecanemab. ⁷

CLARITY AD showed a statistically significant 27% slowing of cognitive decline for lecanemab‐treated versus placebo participants, supporting lecanemab's US Food and Drug Administration approval. ⁷ , ¹² , ¹³ , ¹⁴ However, a supplementary forest plot from CLARITY AD's results paper (Figure S 1‐B in the cited paper's appendix) indicated a statistically significant sex difference in primary clinical effect. ⁷ More specifically, in that plot the point estimate of the effect in females was outside the confidence interval of the male point estimate. The confidence interval for the female subgroup also crossed zero. Overall, the plot shows that males had a statistically significant 43% mean slowing of decline, while females had a non‐significant 12% mean slowing that was significantly different from the effect in males.

Several recent works have therefore concluded lecanemab might not effectively slow decline in females. ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ Others have emphasized the apparent sex difference in effectiveness. ² , ⁵ The CLARITY AD paper's main text did not discuss the sex results; the authors ostensibly concluded there was no significant sex difference, as has another paper, possibly because sex was not significant as a covariate in a subgroup analysis model. ⁷ , ¹⁹ , ²¹ Also, the confidence intervals of the male and female subgroups in the aforementioned forest plot overlapped the 27% cohort mean. ⁷

These interpretations could influence decisions on whether lecanemab is prescribed to female patients. The question remains: How should clinicians interpret the CLARITY AD sex results?

Critically, CLARITY AD was not powered to evaluate efficacy separately for each sex. ⁷ , ¹⁹ A too‐small sample could explain the non‐significance of the effect in females, an effect that nonetheless trended toward favoring lecanemab. ⁷ However, low power does not address lecanemab's smaller clinical effect on females than males, which might be explained by AD‐related sex differences in cognitive decline. Females with prodromal AD have been shown to decline faster than corresponding males. ²² , ²³ , ²⁴ , ²⁵

Alternatively, CLARITY AD's sex difference could be an artifact linked to randomization and participant heterogeneity in disease progression. Studies have shown that patients with similar AD severity at baseline can have very different cognitive decline rates. ²⁶ , ²⁷ , ²⁸ , ²⁹ AD trial simulations have shown that randomization can produce an imbalance of fast‐ and slow‐progressing patients between treatment groups, thus affecting observed drug effect sizes. ³⁰

In this article, we use sex‐specific cognitive decline models and simulations constrained by CLARITY AD's sampling scheme and cohort composition to test the hypothesis that the trial's sex difference could be explained by the phenomena above.

The forest plot described earlier already implies that CLARITY AD's sex difference was probably not an artifact. However, that plot did not report a formal interaction test and does not provide information on whether natural AD sex differences explain the results. By empirically evaluating both possibilities, we aim to conclusively interpret the real trial's sex subgroup results.

2. METHODS

2.1. Experimental reasoning

To test our hypothesis, we ask two basic questions: (1) Could the CLARITY AD sex difference be explained by known pre‐existing sex differences in cognitive decline? (2) Could the sex difference be explained as a random artifactual difference between subgroups?

One question is answered in each of the two experiments. If the answer to either question is “yes,” then our original hypothesis is likely true, implying there was likely no genuine sex difference in lecanemab's clinical effect in CLARITY AD. If the answer to both questions is “no,” then lecanemab may have had a different effect on the trial's males and females.

2.2. Alzheimer's Disease Neuroimaging Initiative

Our analyses used participant data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (https://adni.loni.usc.edu), specifically from the ADNI‐1, ADNI‐GO, ADNI‐2, and ADNI‐3 cohorts. ³¹ , ³² , ³³ , ³⁴ The ADNI was launched in 2003 as a public–private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early AD. For up‐to‐date information, see www.adni‐info.org. ADNI received ethics approval from all participating institutions and received informed consent from all participants.

2.3. Simulation cohort participants

We selected ADNI participants who satisfied CLARITY AD inclusion criteria listed on ClinicalTrials.gov. ³⁵ Data from the selected participants were used in our modeling and simulation experiments. These participants comprise our “simulation cohort.” At baseline, each participant met the clinical criteria for intermediate likelihood AD‐related MCI or probable mild AD dementia, was between 50 and 90 years old, and had Mini‐Mental State Examination (MMSE) scores ≥ 22 and ≤ 30. Participants with MCI had global Clinical Dementia Rating (CDR) scores of 0.5 and CDR Memory Box scores ≥ 0.5. Mild AD dementia participants had global CDR scores of 0.5 or 1.0 and CDR Memory Box scores ≥ 0.5. All participants had a body mass index > 17 and < 35 at baseline. Participants were apolipoprotein E (APOE) ε4 genotyped and were positive for abnormal amyloid levels at baseline. Amyloid status was determined using ADNI's reported baseline PET tracer measures of brain amyloid levels (thresholds: Pittsburgh compound B standardized uptake value ratio [SUVR] > 1.2 or AV45 SUVR > 1.11), ³⁶ , ³⁷ or baseline cerebrospinal fluid measures of amyloid beta 42 (thresholds: Aβ42 ≤ 980 pg/mL). ³⁸ The left panel in Figure 1 shows how the simulation cohort is generated.

FIGURE 1.

Procedure for generating 10,000 clinical trial simulations constrained by the CLARITY AD design. First, ADNI participants who fit CLARITY AD inclusion criteria are selected to form our “simulation cohort.” We then fit the Equation 1 model to data from subsets of that cohort in Experiment 1 (only females and only males) and from the full cohort in Experiment 2. The model longitudinally tracks cohort‐ and participant‐level CDRSBΔbl scores. Next, to simulate a trial, we first randomly resample participants from our simulation cohort and then add noise to the baseline CDRSB score of each sampled participant. This process generates a cohort of synthetic participants—our “synthetic cohort”—for our simulated trial, for example where the sample size is equal to that specified in the CLARITY AD design. We then randomize those participants into “drug” and “placebo” groups. The synthetic cohort data are then input to the Equation 1 model that was originally fit to the simulation cohort. With noise added to the model parameters, Equation 1 is used to calculate CDRSBΔbl scores for each participant at a number of visits that matches the observation scheme of CLARITY AD, up to 18 months. A drug effect is simultaneously injected as a reduction in CDRSBΔbl slope for drug‐treated participants relative to placebo. Once the longitudinal CDRSBΔbl scores are generated (i.e., the trial data simulation is complete), the drug versus placebo group difference in CDRSBΔbl slope is extracted. The extracted slope difference is the observed drug effect size in the simulated trial. This process is repeated 10,000 times. Each simulated trial will have a different observed effect size because of the heterogeneity in natural cognitive decline trajectories in the trial's unique synthetic cohort. Finally, the observed effect sizes across the 10,000 simulations are plotted in a histogram, in which the mean observed drug versus placebo difference in slope is here indicated by the black dashed line. AD, Alzheimer's disease; ADNI, Alzheimer's Disease Neuroimaging Initiative; CDRSB, Clinical Dementia Rating Sum of Boxes; CDRSBΔbl, Clinical Dementia Rating Sum of Boxes change since baseline.

RESEARCH IN CONTEXT

1. Systematic review: The sex subgroup analysis in CLARITY AD, the phase 3 trial of the now US Food and Drug Administration–approved anti‐amyloid drug lecanemab, showed a significant clinical effect in males only. Our literature review (e.g., on PubMed) found that recent work interpreted the sex result as implying that lecanemab has either no or limited effectiveness in females. We found no study had quantitatively and didactically evaluated whether the CLARITY AD result could be explained by pre‐existing sex differences or drug‐independent participant heterogeneity in cognitive decline rate.

2. Interpretation: Our study empirically demonstrates that CLARITY AD's observed sex difference likely cannot be explained by drug‐independent participant heterogeneity in cognitive decline rate, nor by inherent sex differences in decline rate. Lecanemab is likely more effective in males than females but might yet be effective in females.

3. Future directions: Future work should examine possible links between a drug's action mechanism and sex differences in amyloid clearing and clinical efficacy.

2.4. Experiment 1

In Experiment 1 we asked: Given the CLARITY AD trial design and cohort, could the observed sex difference in lecanemab's clinical effect be explained by lecanemab‐independent sex differences in cognitive decline trajectories? We answer this question in two parts.

In Part 1 we use a linear mixed effects model to track cognitive decline separately in the male and female participants in our selected ADNI cohort. Our goal is to determine whether there is an inherent average sex difference in cognitive decline rate.

In Part 2 we use the sex‐specific models from Part 1 to simulate trials adhering to CLARITY AD design parameters in the hypothetical scenarios of including only males or only females. In these scenarios, a simulated drug slows cognitive decline by 27% versus placebo (the overall lecanemab effect in CLARITY AD). ⁷ The sex‐specific placebo decline rate is defined by the estimated slope of the corresponding model from Part 1. If the placebo group's decline rate is sufficiently different between males and females, a sex difference in the observed drug effect size could result. Our goal here is to determine whether drug effect sizes observed in simulated female‐only trials would be different, on average, than effect sizes observed in simulated male‐only trials, despite there being no sex difference in the drug's percentage slowing of decline.

A sufficiently large sex difference in cognitive decline rates and trajectories in Part 1 or in possible observed drug effect sizes in Part 2 might explain the CLARITY AD sex difference result.

2.4.1. Part 1: sex‐specific modeling procedure

We fitted a continuous‐time linear mixed effects model (Equation 1) separately in the male and female subsets of our ADNI simulation cohort (see the left panel of Figure 1). These models longitudinally track group‐ and participant‐level trajectories in CDR Sum of Boxes change since baseline (CDRSBΔbl). Increasing CDRSB scores indicate worsening impairment. ³⁹ All our analyses were done in R version 4.3.2 in macOS 14.2.1. ⁴⁰ To fit the model, we used the function lmer in package lme4. ⁴¹

Our model is a continuous‐time analog of the categorical‐time primary analysis mixed model for repeated measures (MMRM) from CLARITY AD. ⁷ An MMRM would let us evaluate sex differences in the magnitude of cognitive decline at specific time points (e.g., relative to baseline) and could approximately replicate the CLARITY AD analysis. We selected a different model for two main reasons. First, our approach models change continuously, which lets us more intuitively communicate any sex difference in the overall trend and rate of cognitive decline compared to an MMRM. Second, continuously modeling the actual time of data acquisition in ADNI better leverages both time and cognitive test scores to fit trajectories for individual participants, which is critical for simulating clinical trials in Experiment 1 Part 2 and Experiment 2 (see the corresponding Methods for details). Equation 1 provides our model's syntax.

$$\begin{array}{ccc}& & CDRSB\_change\_since\_baseline\sim 0+CDRSB\_baseline\_median\_centered\hfill \\ & & +Diagnosis\_AD+Years\_since\_baseline+APOE4\_status\hfill \\ & & +CDRSB\_baseline\_median\_centered:Years\_since\_baseline\hfill \\ & & +(0+Years\_since\_baseline|participant\_ID)\hfill \end{array}$$ (1)

The dependent variable is $CDRSB\_change\_since\_baseline$, which signifies the magnitude of a participant's change in CDRSB score relative to their score at baseline. This variable has the same units as the corresponding variable in CLARITY AD's MMRM.

The independent variables correspond to those in CLARITY AD's MMRM, with some key differences. ⁷ Median‐centered baseline CDRSB scores ($CDRSB\_baseline\_median\_centered$) are used instead of baseline CDRSB scores to help ensure reliable parameter estimates. Median centering here is done by subtracting the median value of all baseline CDRSB scores across our selected cohort from each participant's baseline CDRSB. $Years\_since\_baseline$ is the continuous time variable and indicates number of years elapsed since a participant's baseline visit in ADNI. In the fitted model, the coefficient of this variable signifies the cohort‐average rate of cognitive decline (units of change in CDRSBΔbl per year). This time variable is used in place of the categorical “Visit” term in CLARITY AD's MMRM. We also include an interaction term for median‐centered baseline CDRSB scores with years since baseline ($CDRSB\_baseline\_median\_centered:Years\_since\_baseline$). This term corresponds to the CLARITY AD model's interaction term of baseline CDRSB with categorical “Visit.” Our model also includes two terms ported directly from CLARITY AD's MMRM: AD diagnosis ($Diagnosis\_AD$, AD = 1 or MCI = 0), and APOE ε4 status ($APOE4\_status$, homozygote or heterozygote = 1 or non‐carrier = 0).

There are four terms in CLARITY AD's model that we did not include here: geographic region (defined in the trial as North America, Europe, and Asia‐Pacific), trial group (i.e., lecanemab group or placebo group), trial group interaction with time (time being categorical “Visit” in CLARITY AD's model), and baseline use of medication for AD symptoms. ADNI collected data at different sites only in North America, so we are unable to replicate CLARITY AD's region term. There is no monolithic group‐partitioned drug effect like lecanemab's in ADNI, so we do not include trial group here. However, trial group does factor into the simulations in Experiment 1 Part 2 and Experiment 2 (see the corresponding Methods). Information on the use of medication for AD symptoms was not available in all our selected ADNI participants so we did not include a corresponding term in our model.

Another key difference of our model compared to CLARITY AD's is that ours encodes random effects by participant for “$Years\_since\_baseline$.” This addition lets our model estimate each participant's cognitive decline rate, quantified as a deviation from the cohort's average decline rate (same units as described above for the coefficient of “$Years\_since\_baseline$”).

No intercepts are encoded for the model because all participants’ CDRSBΔbl scores ($CDRSB\_change\_since\_baseline)$ at baseline are zero, that is, there is no change since baseline at baseline.

To complete Part 1 of Experiment 1, we compared the male and female models’ estimated coefficients on the “$Years\_since\_baseline$” term and their standard errors to evaluate sex differences in cognitive decline rate.

For Part 2, we used the sex‐specific models separately to generate synthetic trial data (i.e., synthetic $CDRSB\_change\_since\_baseline$ scores over time) in male‐only and female‐only simulated trials. This process is illustrated in the middle and right panels of Figure 1 and is described below.

2.4.2. Part 2: sex‐specific trial simulation procedure

Using our female‐only trial simulation as the example here, we first randomly sampled female participants from our simulation cohort, with replacement. This resampling procedure generated baseline data for 1800 synthetic female participants, that is, the approximate full sample size targeted in the CLARITY AD design. The resampling preserved the approximate baseline AD stage distribution of the real trial, in which ≈ 60% of participants had MCI and 40% had mild AD dementia at baseline. ⁷ For each sampled participant, random noise corresponding to measurement error was added to their baseline CDRSB score. ⁴² Noise was also added to the cohort‐level and participant‐specific model parameters to account for uncertainty and to further differentiate synthetic participants in our resulting “synthetic cohort.”

Next, we randomized these participants 1:1 to “drug” and “placebo” groups. We then used the participants’ baseline data and the estimated parameters from our female‐only model to calculate CDRSBΔbl scores (i.e., $CDRSB\_change\_since\_baseline$ values) for each participant at seven visits up to 18 months of follow‐up. Random noise was added to each participant's calculated CDRSBΔbl score at each visit to incorporate uncertainty and measurement error. Noise values were drawn from a zero mean normal distribution with a standard deviation equal to that of the observation‐level residuals in our female‐only simulation cohort model. Calculated CDRSBΔbl scores were then rounded to correspond to standard CDRSB 0.5 increments (e.g., 1.1 rounded to 1.0, and 1.4 rounded to 1.5). ³⁹ Visits were defined to occur every 3 months, and each participant's specific visit time values had random noise added to simulate an approximate ± 1‐month window around each visit. This observation scheme approximates that of CLARITY AD. Note that ADNI did not observe CDRSB every 3 months. ³¹ , ³² , ³³ , ³⁴ By adding noise values in the manner described above to the model parameters and to each simulated CDRSBΔbl score, we aim to account for the uncertainty associated with calculating CDRSBΔbl scores for the defined observation scheme.

A drug effect was simultaneously injected as a 27% reduction in the linear slope of drug‐treated synthetic participants’ CDRSBΔbl trajectories compared to the placebo participants. The slope for the placebo group is equal to the β (i.e., the coefficient) of the “$Years\_since\_baseline$” term in the Equation 1 model fit to our female‐only simulation cohort. In this way, our female‐only trial simulation model encodes a female‐specific rate of cognitive decline for the placebo group, and thus a female‐specific decline rate for the drug group. We also encoded 20% participant dropout by the trial's final observation, as was assumed in the real CLARITY AD sample size calculation. ⁷

This overall procedure generates a simulated trial constrained by the CLARITY AD parameters specified in its analysis plan, as described in the main results paper. ⁷ The process was repeated 10,000 times to simulate 10,000 female‐only trials matching the CLARITY AD design parameters. Each simulation had the same cohort‐level trend but a different sampling of participant‐level trajectories, thus accounting for randomization and inter‐participant heterogeneity in cognitive decline within and across simulations. ³⁰ Next, we extracted the drug versus placebo CDRSBΔbl linear slope difference (i.e., the observed drug effect size) from each simulation using the Equation 1 model, but here it contained an additional interaction term of trial group (drug = 1, placebo = 0) with $Years\_since\_baseline$. The estimated coefficient of this new term represents the observed drug effect size in the simulated trial. While a standalone trial group term was included in CLARITY AD's primary analysis model, we do not include that term here because our simulated drug effect is encoded only as a slope‐reducing effect, that is, a cognitive decline rate–reducing effect as reported in CLARITY AD's primary analysis. We do not encode any other offset in effect between trial groups. Overall, this methodology is conceptually analogous to simultaneously running 10,000 identically designed female‐only trials in the real world.

We also applied this procedure using the male‐only data and model to generate 10,000 male‐only trial simulations. Overall, we obtain a set of 10,000 observed effect sizes for the female‐only simulated trials, and 10,000 for the male‐only trials.

To finish Experiment 1 Part 2, we evaluated differences between our female and male effect size distributions using a t test to compare means and a Bartlett test to compare variances. ⁴³ , ⁴⁴ We also calculated the percent difference in those means.

While the 10,000 female‐only and 10,000 male‐only simulated trials all had the same cohort‐level drug effect, we hypothesize that the mean and variance of the effect size distributions would be different between the sexes. A difference in means could result because the placebo group's cognitive decline rate (parameterized as the β for the “$Years\_since\_baseline$” term from Equation 1) differed between female‐only and male‐only simulations. A difference in variance could result because the female‐only and male‐only cohorts contained different cognitive decline trajectories. Other parameters from Equation 1 (e.g., the estimated β values and standard errors) also differed between the female‐only and male‐only models and trial simulators.

Our analysis approach could also be applied to CLARITY AD's secondary outcome metrics (e.g., Alzheimer's Disease Assessment Scale–Cognitive subscale [ADAS‐Cog14]). We prioritized CDRSB because of its primary outcome status in CLARITY AD, which made CDRSB the focus of recent works that highlighted a sex difference in lecanemab's effectiveness or concluded the drug might be ineffective in females. ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ We aim to help resolve this ambiguity. CDRSB was also present in all our ADNI datasets, giving us the maximum sample possible. ADAS‐Cog14, for example, was not available in every ADNI dataset we used. ³¹ , ³² , ³³ , ³⁴

2.5. Experiment 2

In Experiment 2 we asked: Could CLARITY AD's observed sex difference have been a random artifactual difference between the male and female subgroups? An artifact here represents an observed sex difference in drug effect when there is no genuine sex difference in the effect (i.e., a Type 1 error).

We answer this question using simulations, posing it technically as: What is the probability of observing a ≥ 31% difference in a drug's effect between two equally sized subgroups containing participants selected totally at random? This 31% was the difference between the male and female effects from lecanemab in CLARITY AD. If the probability is sufficiently high (e.g., > 0.05), then the sex difference in CLARITY AD might have been an artifact.

2.5.1. Trial simulation procedure

Here we simulate trials using the same CLARITY AD design parameters as in Experiment 1. We use the same model as in Equation 1, but here we fit the model to data from all participants in our ADNI simulation cohort. The same trial simulation procedure as in Experiment 1 is then applied, but here using the pooled male and female data. However, this time we randomize 1800 synthetic participants 1:1 to “Subgroup 1” and “Subgroup 2,” signifying two subgroups of 900 participants selected totally at random with no consideration of sex. This subgroup sample size is approximately equal to that of the male and female subgroups in the CLARITY AD design. ⁷

After running 10,000 simulations, we extracted the Subgroup 1 versus Subgroup 2 CDRSBΔbl linear slope difference (i.e., the observed difference in drug effect between the subgroups) from each simulation. Next, we calculated the proportion of simulations with a between‐subgroup difference ≥ 31%, yielding the probability of at least that difference arising as a random artifact.

3. RESULTS

3.1. Participant demographics

Of the 2420 ADNI participants available, 644 met the specified CLARITY AD inclusion criteria, with 264 (41%) being female and 380 (59%) being male. The mean baseline ages (± standard deviations) for females and males, respectively, were 71.6 ± 7.3 years and 73.9 ± 6.8 years. The mean number of years of education was 15.3 ± 2.7 for females and 16.4 ± 2.8 for males. One hundred eighty‐three females (69%) and 250 males (66%) were APOE ε4 heterozygotic or homozygotic. At baseline, 76 females (29%) and 102 males (27%) were diagnosed with mild AD dementia, with the remaining participants having a baseline diagnosis of AD‐related MCI. See Table 1 for the demographic summary.

TABLE 1.

Demographics of our simulation cohort participants selected from ADNI.

	Females (n = 264)	Males (n = 380)
Age (mean years ± SD)	71.6 ± 7.3	73.9 ± 6.8
Education (mean years ± SD)	15.3 ± 2.7	16.4 ± 2.8
APOE ε4 positive	183 (69%)	250 (66%)
Mild AD dementia	76 (29%)	102 (27%)
Mild cognitive impairment	188 (71%)	278 (73%)

Abbreviations: AD, Alzheimer's disease; ADNI, Alzheimer's Disease Neuroimaging Initiative; APOE, apolipoprotein E; SD, standard deviation.

3.2. Experiment 1 results

3.2.1. Part 1: sex‐specific modeling results

Females and males who fit CLARITY AD inclusion criteria had similar CDRSBΔbl trajectories estimated by the Equation 1 model. We used the Nakagawa conditional R ² for mixed models as an absolute value goodness‐of‐fit measure, yielding 0.931 and 0.928 for the male‐only and female‐only models, respectively, indicating well‐fitting models. ⁴⁵ , ⁴⁶ Key parameters for the male‐only and female‐only models are summarized in Table 2A and B, respectively. Figure 2A shows estimated cognitive decline trends by sex, as well as real trajectories for individual participants. There was no statistically significant sex difference in the estimated cognitive decline rate between the male and female models, as the confidence intervals on the “$Years\_since\_baseline$” β value of each sex overlaps the opposing sex β value on this term (see Table 2A,B).

TABLE 2.

Key estimated parameters from the Equation 1 model fit to data from (A) only male participants, (B) only female participants, and (C) male and female participants together in the simulation cohort.

(A) Male‐only model Nakagawa conditional R 2 = 0.931	𝛽	Standard error	t value	p value
Median‐centered baseline CDRSB	−0.08145	0.03508	−2.322	0.020350
Baseline AD diagnosis status	0.33407	0.11964	2.792	0.005285
Years since baseline	0.77923	0.06486	12.014	< 2e−16
APOE ε4 status	−0.18959	0.04935	−3.841	0.000127
Interaction of median‐centered baseline CDRSB with years since baseline	0.31434	0.04464	7.041	7.6e−12

(B) Female‐only model Nakagawa conditional R 2 = 0.928	𝛽	Standard error	t value	p value
Median‐centered baseline CDRSB	−0.16455	0.04115	−3.999	6.76e−05
Baseline AD diagnosis status	0.53409	0.14458	3.694	0.00023
Years since baseline	0.84646	0.08339	10.151	< 2e−16
APOE ε4 status	−0.32775	0.05849	−5.604	2.64e−08
Interaction of median‐centered baseline CDRSB with years since baseline	0.35413	0.05168	6.852	4.70e−11

(C) Male and female model Nakagawa conditional R 2 = 0.930	𝛽	Standard error	t value	p value
Median‐centered baseline CDRSB	−0.11808	0.02668	−4.426	9.96e−06
Baseline AD diagnosis status	0.41625	0.09223	4.513	6.62e−06
Years since baseline	0.80663	0.05113	15.777	< 2e−16
APOE ε4 status	−0.24694	0.03774	−6.543	7.12e−11
Interaction of median‐centered baseline CDRSB with years since baseline	0.33388	0.03365	9.923	< 2e−16

Abbreviations: AD, Alzheimer's disease; APOE, apolipoprotein E; CDRSB, Clinical Dementia Rating Sum of Boxes.

FIGURE 2.

CDRSBΔbl trajectories for our simulation cohort participants selected from ADNI. Scores are tracked over number of years elapsed since a participant's baseline ADNI observation. A, Real individual trajectories for all simulation cohort participants, with overlaid female‐ and male‐specific group‐level trends estimated by Equation 1 indicated in red and blue, respectively, with confidence intervals. These two trends nearly overlap for the 18‐month period that corresponds to the CLARITY AD trial duration. There is no statistically significant sex difference in the estimated cognitive decline rate, as the confidence intervals on the slope values overlap between the two models (see Table 2). The models describing these trends were used in generating the Experiment 1 simulations. B, Real individual trajectories for all simulation cohort participants with the cohort‐level trend line estimated by Equation 1 indicated in green, with confidence interval. The model describing this trend was used in generating the Experiment 2 simulations. ADNI, Alzheimer's Disease Neuroimaging Initiative; CDRSBΔbl, Clinical Dementia Rating Sum of Boxes change since baseline

3.2.2. Part 2: sex‐specific trial simulation results

Figure 3 shows histograms of simulated CDRSBΔbl slope differences (i.e., differences in cognitive decline rate) between drug and placebo groups, taken from within 10,000 male‐only and 10,000 female‐only simulated trials adhering to the CLARITY AD trial design parameters. The horizontal axis of the histogram signifies values of the difference in cognitive decline rate between a trial's drug and placebo groups. Each value is obtained from within one simulated trial instance and equals the drug‐treated group's slope on its CDRSBΔbl trend minus the placebo group's slope on its CDRSBΔbl trend. The female‐only simulations and the male‐only simulations each have their own histogram (red and blue, respectively, in Figure 3). Each histogram thus illustrates the distribution of observed drug effect sizes across the corresponding 10,000 sex‐specific simulations.

FIGURE 3.

Histograms of simulated CDRSBΔbl slope differences for drug‐treated versus placebo groups (i.e., observed drug effect sizes) for simulated trials adhering to the CLARITY AD design parameters, but including only males (blue) or only females (red). Each histogram represents 10,000 simulations. The mean value for each distribution is indicated by a color‐coded dashed line. The small difference in means is statistically significant, as is the difference in variances. CDRSBΔbl, Clinical Dementia Rating Sum of Boxes change since baseline

Every simulated trial had the same cohort‐level drug effect encoded as a 27% reduction in CDRSBΔbl slope for drug versus placebo participants. However, the decline rate for the placebo group differed between the female‐only and male‐only simulations. Within each trial, the placebo decline rate was defined as the β for the “$Years\_since\_baseline$” term in Table 2; the male‐only and female‐only models each have their own value for this β. Thus, as the drug effect in each trial was encoded as a percent reduction in the placebo decline rate, the 10,000 male‐only and 10,000 female‐only simulated trials had a different value on average for the observed drug effect. This on‐average difference explains the offset positioning of the histograms in Figure 3.

A Bartlett test of homogeneity of variances showed that the distributions’ variances were significantly different (K ² = 7.4819, P = 0.006232). ⁴³ , ⁴⁴ A two‐sample t test showed a statistically significant difference in means for the distributions (t = 21.306, P < 2.2e−16). The mean effect sizes from the male and female simulations are indicated in Figure 3. The means are µ _male = −0.211 and µ _female = −0.229, and the standard deviations are σ _male = 0.059 and σ _female = 0.061. The mean from the male‐only simulations is ≈ 7.9% less than the mean from the female‐only simulations.

This difference in means is linked directly to the statistically non‐significant sex difference in cognitive decline rate in our ADNI simulation cohort, that is, the difference between the β values for “$Years\_since\_baseline$” in Table 2A and B. Concretely, the real male participants in our selected ADNI cohort had a numerically smaller cognitive decline rate than the real female participants. In a sex‐specific simulated trial, the sex‐specific decline rate from the corresponding real ADNI participants defined the placebo group's decline rate: The male trial simulations used the male ADNI decline rate for the placebo group, and similarly for the female simulated trials. In all sex‐specific trial simulations (male and female), the drug effect was defined as a 27% reduction in cognitive decline rate versus the placebo group's decline rate. The observed male drug effect size was smaller on average than the female one because the male placebo decline rate was numerically smaller than the female placebo decline rate.

When using only 1000 simulations of the male or female trial scenarios, the statistically significant difference in means persisted (t = 5.8396, P = 6.097e−09), but the t value decreased and the P value increased compared to using 10,000 simulations. The difference in statistical significance for 1000 versus 10,000 simulations is linked primarily to statistical power: The more sex‐specific trials we simulate (in which there is a true difference in mean drug effect between the male and female simulations), the higher the probability of obtaining a statistically significant estimate for that difference in means. The implication here is that even with a reduced number of sex‐specific trials simulated—1000 versus 10,000—we still obtained a statistically significant estimate for the difference in mean observed drug effect size between the male‐only and female‐only simulations. Thus, the statistically non‐significant sex difference in natural cognitive decline rate (summarized in Table 2A,B) was large enough to produce a sex difference in the mean simulated drug effect when using only 1000 simulations, a scenario in which statistical power to detect the difference in means was much lower compared to using 10,000 simulations.

3.3. Experiment 2 results

Figure 4 shows a histogram of 10,000 simulated CDRSBΔbl slope difference values, each signifying the difference in drug effect size between two subgroups of 900 randomly selected participants from within one simulated trial. Male and female participants were included at random in each simulation. The Equation 1 model, here fit to data from all our simulation cohort participants, was used to generate the simulations. Table 2 lists key model parameters, where a Nakagawa conditional R ² of 0.930 indicates a well‐fitting model. Figure 2B shows the estimated overall cognitive decline trend and individual trajectories for all real participants in the simulation cohort.

FIGURE 4.

Histogram of differences in CDRSBΔbl slope (here signifying differences in drug effect size) between two subgroups of randomly selected participants in 10,000 trial simulations constrained by the CLARITY AD design. The red dashed line indicates the value corresponding to the 31% difference between sex subgroups reported in CLARITY AD. The black dashed line corresponds to the preloaded known sex difference in drug effect from Experiment 1. Only 12 of 10,000 simulations (indicated in purple to the left of the red line) had a subgroup difference ≥ 31%. Even when a large difference is preloaded between subgroups, there is an extremely low probability of randomly observing a ≥ 31% difference in drug effect between subgroups. CDRSBΔbl, Clinical Dementia Rating Sum‐of‐Boxes change since baseline

The red vertical line in Figure 4 indicates the difference in drug effect size between subgroups that corresponds to the CLARITY AD–reported 31% difference between its male and female subgroups. A ≥ 31% difference here implies that one subgroup's participants have at least a 31% greater reduction in cognitive decline rate compared to the other subgroup.

Additionally, in each simulation, we encoded a 7.9% between‐subgroup difference in drug effect. With this bias, one subgroup always declines 7.9% slower than the other, resulting in a corresponding on‐average difference in decline rate between subgroups across our 10,000 simulations (indicated by the black line in Figure 4). This 7.9% equals the percent difference in mean drug effect size between males and females in our Experiment 1. This bias effectively preloads the known sex difference in drug effect and thus biases the simulation in favor of CLARITY AD. Any further difference between subgroups is attributable to participant heterogeneity and randomization.

Only 12 of the 10,000 simulated trials had a subgroup difference in drug effect ≥ 31%, even with our known sex difference in drug effect preloaded. This result signifies a 0.0012 probability of observing a drug effect difference ≥ 31% between two subgroups, even when there is a known difference of 7.9% between the subgroups.

4. DISCUSSION

We examined whether the sex difference in lecanemab's clinical effect observed in CLARITY AD could be explained by inherent sex differences in cognitive decline or as an artifact linked to participant heterogeneity and randomization. Our selection of ADNI participants who met CLARITY AD's inclusion criteria resulted in a female cohort that was slightly younger and less educated (age: 71.6 ± 7.3 years, education: 15.3 ± 2.7 years) than the males (age: 73.9 ± 6.8 years, education: 16.4 ± 2.8 years); the groups had similar proportions of APOE ε4–positive individuals and a similar split of MCI and mild AD dementia participants.

The yearly change in CDRSB, estimated in Experiment 1 Part 1 from the β values of the sex‐specific models (Equation 1), was not statistically significantly different between males (β _male = 0.77923 ± 0.06486) and females (β _female = 0.84646 ± 0.08339). In other words, even though the graph in Figure 2A shows a slight sex‐related difference, the confidence intervals (in gray) overlap for the 10+ year period fitted, and the two regression lines overlap almost completely in the 18‐month period corresponding to the CLARITY AD trial duration.

Nonetheless, Experiment 1 Part 2 showed a statistically significant difference in mean observed drug effect size between male‐only and female‐only simulated trials constrained by CLARITY AD design parameters, in which we encoded the same drug effect in all simulations as a 27% reduction in decline rate for drug‐treated versus placebo participants. Males had a mean observed effect that was ≈ 7.9% smaller than females. This sex difference in mean effect occurred because the simulated male placebo groups had a smaller rate of cognitive decline than the simulated female placebo groups. Overall, this 7.9% sex difference does not explain the 31% difference in drug effect between CLARITY AD's male and female participants. Moreover, CLARITY AD reported a larger effect for males than females, while our outcome was the opposite. Given our results and the magnitude and direction of CLARITY AD's sex difference, it is unlikely that the trial's males and females experienced the same drug effect.

The answer to our Experiment 1 question is therefore: No, CLARITY AD's sex difference cannot be explained by lecanemab‐independent sex differences in cognitive decline. Our simulations provide empirical justification.

Experiment 2 showed that a ≥ 31% difference in drug effect between subgroups is extremely unlikely to occur as a random artifact, even when a known difference between subgroups is present. Based on our simulations, the estimated probability of that occurrence is 0.0012. Experiment 2 produced the same conclusion as an expert reading of the key CLARITY AD forest plot (Figure S 1‐B in the cited paper's appendix), in which the confidence intervals on the male and female mean effects do not overlap the mean effect of the opposing sex. ⁷ The answer to our Experiment 2 question is thus: No, the CLARITY AD sex difference cannot be explained as an artifactual difference between the male and female subgroups.

These results imply that the CLARITY AD sex difference is unlikely to be explained by the phenomena we investigated. CLARITY AD also showed statistically significant larger effects for lecanemab in males than in females in two of three non‐primary clinical endpoints (AD Composite Score and Alzheimer's Disease Cooperative Study MCI Activities of Daily Living). ⁷ The point estimate of lecanemab's effect was numerically larger in males than females for the third non‐primary endpoint, ADAS‐Cog14, but the difference was not statistically significant. ⁷ However, the sex subgroup analysis of four quality‐of‐life (QOL) metrics showed no statistically significant sex difference in QOL change since baseline at 18 months. ⁴⁷ Nonetheless, point estimates on three of the metrics (Zarit Burden Interview total score and QOL in AD total score [A] by subject and [B] by subject by proxy) trended toward larger benefit in males than females. ⁴⁸ , ⁴⁹

Given these outcomes, CLARITY AD likely observed a genuine sex difference in lecanemab's clinical effect. Further research is needed to determine whether such a difference could be related to biological or gender differences, such as in hormone profiles, cognitive reserve, and education. ² Our study points to a possible sex difference in the clinical efficacy of amyloid‐targeting drugs, specifically for lecanemab, and possibly for those with different mechanisms of action such as aducanumab and donanemab. ⁶ , ⁷ , ⁸

In aducanumab's phase 3 trial EMERGE, a statistically significant primary clinical effect was reported only for the male subgroup in the main results paper's supplementary Figure 3‐A forest plot (see citation). ⁶ However, there was no significant sex difference in that effect. Like lecanemab, aducanumab trended toward effectiveness in females, and the non‐significance of the effect could be explained by low power. Similarly, no significant sex difference in primary efficacy was observed in donanemab's positive phase 3 trial TRAILBLAZER‐ALZ 2. ⁸ However, the eFigure 9‐B forest plot in the results paper's supplement indicates a statistically non‐significant effect in males only (see citation). ⁸ This non‐significance could again be due to low power.

These outcomes suggest that CLARITY AD's sex difference could be linked to lecanemab's mechanism of action. Unfortunately, to the best of our knowledge, sex‐disaggregated data on amyloid clearance are not publicly available for the trials discussed. Sex analyses of CLARITY AD's amyloid PET data might reveal whether the sex difference in efficacy is accompanied by corresponding differences in amyloid clearing. A sex difference in the clinical impact of removing amyloid would fit conceptually with previously observed modulating effects of sex on the association between AD pathology and clinical progression. For example, amyloid plaques and neurofibrillary tangles have been observed to manifest clinically as dementia at higher rates in females than in males. ²⁵ Alternatively, a sex difference in amyloid clearance alone could explain CLARITY AD's observed sex difference in lecanemab's efficacy. A comparison of CLARITY AD's sex‐disaggregated PET data with corresponding data from the phase 3 aducanumab and donanemab trials could highlight links between a drug's action mechanism and any sex differences in amyloid clearing and clinical efficacy.

While CLARITY AD had near‐equal sex representation in its lecanemab and placebo groups, future trials might benefit from recruiting larger cohorts while preserving sex parity. Larger sample sizes could ensure adequate power to definitively observe sex differences, including in preclinical trial phases and drug development research. ²² Stratification within sex subgroups, such as by menopausal state or gender‐biased risk factors (e.g., lower education in women than men) might help disentangle sex and gender contributions to any male/female difference in a drug's clinical effect. ⁵⁰ , ⁵¹

CLARITY AD also reported differences in drug effect for older versus younger participants, and between APOE ε4 non‐carriers or heterozygotes versus homozygotes. ⁷ Well‐powered subgroup analyses might thus be needed for demographic categories in addition to sex. Cohort enrichment techniques (e.g., subject selection via cognitive decline prediction) could enhance the power to detect effects within subgroups without requiring more participants. ⁵² , ⁵³ Advanced trial designs such as those relying on digital twins might also require fewer participants than conventional randomized placebo‐controlled trials. ⁵⁴

For now, however, subgroup analyses in AD trials should be reported and interpreted cautiously. As evidenced by recent works examining CLARITY AD's sex difference, the current style of reporting AD trial subgroup results in forest plots might lead to misinterpretations of statistically non‐significant effects in low‐powered subgroups. The CLARITY AD paper's authors themselves stated that the trial was not powered to evaluate subgroups separately. ¹⁹ The phase 3 aducanumab and donanemab trials could also suffer from such misinterpretations. Unless a trial is powered to evaluate efficacy within each subgroup, single P values (and/or confidence intervals) in subgroup analyses should only be used to indicate whether treatment was significantly more effective in one subgroup than another. Trials should still be powered to evaluate such differences.

Overall, our analysis has some limitations. We used a linear model to generate our simulations, but AD progresses non‐linearly over the full disease timeline. ⁵⁵ However, our simulations cover trajectories over 18 months only (the CLARITY AD trial duration), so linear assumptions are reasonable. Linear mixed effects models also capture inter‐participant heterogeneity in cognitive decline around a cohort‐level average trend, even across participants with MCI or mild dementia. ⁵⁶ Future analyses might use non‐linear models, including ones that make no assumptions about disease trajectories when analyzing trial data. ⁵⁶

We also do not include placebo effects, so our simulations essentially compare a drug effect to ADNI's standard of care. ADNI participants were recruited across Canada and the United States, where care standards can differ between and within the countries. ⁵⁷ , ⁵⁸ Also, ADNI participants are predominantly highly educated, White, and male. ⁵⁹ , ⁶⁰ , ⁶¹ CLARITY AD was more diverse, with participants having been recruited from across three continental regions (North America, Europe, and Asia‐Pacific). ⁷ Future work might use open datasets from diverse past international trials (e.g., from the Critical Path for Alzheimer's Disease). ⁶²

To conclude, our results combine with the published CLARITY AD sex outcome to suggest that lecanemab has lower clinical benefits in females than in males. Results from CLARITY AD's open‐label extension might clarify the longer term impact of lecanemab in females. Also, certain types of amyloid‐targeting drugs might function differently in females and males. Research into possible mechanisms could be accelerated by drug developers sharing recent AD trial data. ⁶ , ⁷ , ⁸

CONFLICT OF INTEREST STATEMENT

Daniel Andrews declares no conflicts of interest. Dr. Simon Ducharme has received grants or contracts from the Alzheimer's Drug Discovery Foundation, the Canadian Institutes for Health Research (CIHR), the Fonds de recherche du Québec, Novo Nordisk, Biogen, Janssen, Alnylam, and Innodem Neurosciences. Dr. Ducharme has received consulting fees from Eisai, QuRALIS, and Eli Lilly. Dr. Ducharme has received payment or honoraria from Eisai for presentations, manuscript writing, or educational events. Dr. Ducharme has participated on a data safety monitoring or advisory board for IntelGenX and Aviado Bio. Dr. Ducharme is co‐founder and holds stock or stock options in AFX Medical Inc. Dr. Howard Chertkow has received research grants from the following organizations: CIHR, Alzheimer's Society of Canada, BrightFocus Ltd. (grant number A2022046S), National Institutes of Health (grant number 1R01AG075111‐01A1). Dr. Chertkow has received industry‐associated grants from the following organizations: IntelGenx Corp., Alector Inc., Eli Lilly & Co., Biogen MA Inc., and Hoffman‐La Roche. Dr. Chertkow has participated on National Advisory boards for Eisai, Biogen, and Lilly. Dr. Chertkow is the Scientific Director of the Canadian Consortium on Neurodegeneration in Aging. Dr. Chertkow heads up a clinical trials unit and is site investigator for phase 2 and 3 international drug studies sponsored by Roche, Biogen, Eisai, BMS, Alector, IntelGenX, and Anavex. Dr. Maria Pia Sormani has received consulting fees from Biogen, Merck, Sanofi, Roche, Novartis, Alexion, and Immunic. Dr. Sormani has participated on a data safety monitoring or advisory board for Novartis and Sanofi. Dr. D. Louis Collins declares no conflicts of interest. Author disclosures are available in the supporting information.

CONSENT STATEMENT

ADNI received ethics approval from all participating institutions and received informed consent from all participants.

DIVERSITY STATEMENT

Our study examines a possible sex difference in the clinical efficacy of lecanemab, an anti‐amyloid drug designed to slow Alzheimer's disease progression. Sex is thus a core consideration in all parts of this work. Participant characteristics such as race or gender (as different than sex) were not considered as disaggregated factors due to data availability limitations. However, our discussion of the limitations of standard trial subgroup analyses also applies to analyses of demographic groupings other than sex. Future work with more diverse datasets than ADNI will explore possible differences in drug efficacy between additional subgroups.

Supporting information

Supporting Information

Andrews D, Ducharme S, Chertkow H, Sormani MP, Collins DL. The higher benefit of lecanemab in males compared to females in CLARITY AD is probably due to a real sex effect. Alzheimer's Dement. 2025;21:e14467. 10.1002/alz.14467