Anxiety: Effects & Evidence Base — what Meta-analyses support

Introduction

Anxiety isn’t a single symptom with one cause; it’s an umbrella term for different disorder types, severity levels, and measurement endpoints. As a result, study results are often only partially directly comparable—and meta-analyses show trends rather than a perfect “one intervention for everything.” In this article, I systematically categorize evidence for anxiety: from lifestyle to special topics like ketogenic diets or GLP-1 receptor agonists.

Why “anxiety” isn’t the same as “anxiety”: Target, endpoints, and study design

Short answer: Anxiety research often mixes different disorders and outcomes (symptoms, functional ability, sleep), making effects across studies only limitedly comparable. Meta-analyses combine results, but because of differences in measurement tools, populations, and study duration, heterogeneity is common—and the findings must therefore be interpreted in context.

Clinically, anxiety as a term is “broad”: it can appear as an anxiety disorder (e.g., generalized anxiety, specific fears, anxiety disorders within the context of other conditions), as an anxiety state with stress-related peaks, or as part of illnesses that also affect sleep and mood. When studies use “anxiety” as an outcome, it can therefore mean very different things: symptom scales, spontaneous restlessness, behavioral tests—or even indirect markers such as sleep quality.

These differences show up directly in study design and interpretation. Meta-analyses do combine studies, but they can only aggregate what is methodologically comparable in the included work. This is especially relevant because anxiety studies are often conducted in populations where anxiety is linked with other issues—e.g., in dementia or in PCOS (polycystic ovary syndrome) (Nimmons et al., 2024, PMID 38097097; Tan et al., 2026, PMID 41664652). In such cases, “anxiety” may be partly a secondary outcome of another primary effect (e.g., mood improvement, sleep improvement, cognitive/social effects).

Another key point for self-decisions is: Which benefit was measured? Was it about:

• Symptom reduction (scale scores),
• functional outcomes (e.g., participation, daily-life ability),
• or adjacent outcomes like sleep quality?

Because anxiety and sleep are tightly coupled, an intervention (e.g., a physically structured practice) may primarily improve sleep and only secondarily improve anxiety. Therefore, this overview distinguishes between evidence for anxiety itself and evidence for related outcomes—so you don’t compare apples to oranges when evaluating effect sizes or safety risks.

Lifestyle interventions are also heterogeneous: “More movement” isn’t the same as “training X, Y, or Z.” Duration, intensity, frequency, and adherence vary widely. With Tai Chi, for example, older-adult data show effects, but what counts as an “optimal dose range” depends on the evidence base (Chen et al., 2025, PMID 40625025). This is a good example of how, even for the same intervention, the study framework determines what is ultimately measured.

Evidence hierarchy in anxiety: RCTs, meta-analyses, observational data, and animal studies

Short answer: For effectiveness questions, RCTs and the meta-analyses derived from them are the best foundation because they support causal inference better than observational data. Animal models can support mechanisms, but they aren’t automatically transferable to humans—especially for psychological endpoints.

In the evidence hierarchy, a meta-analysis (e.g., based on randomized trials) usually ranks above individual studies because it statistically combines effect sizes, reducing random fluctuations. This is particularly relevant for interventions where you expect a clear pre-to-post change. Examples include meta-analyses on Tai Chi (Chen et al., 2025, PMID 40625025) or on treatment strategies for anxiety in specific populations (Nimmons et al., 2024, PMID 38097097; Tan et al., 2026, PMID 41664652).

What matters here: “higher evidence” doesn’t automatically mean “equal transferability.” Network meta-analyses can rank interventions, but the strength of the conclusion depends on how well the studies are comparable and how solid the initial evidence base is (Tan et al., 2026, PMID 41664652). If populations, endpoints, or co-treatments differ substantially, even a meta-analysis can only partially “disentangle” the signal.

Besides effectiveness, safety always matters in anxiety. This is often underestimated because some anxiety studies focus less on side-effect reporting. The importance of clean safety data is illustrated by how systematic reviews evaluate RCTs (even when the specific example is about daytime sleepiness/sleep): when classifying such data, it’s crucial which adverse events were actually reported and extracted in RCTs (Jalal et al., 2026, PMID 41324388). For anxiety, the same logic applies: if safety endpoints are reported incompletely, your safety evaluation is limited—even if efficacy on scales looks favorable.

For medications, there’s also a key dividing line between anxiety-related endpoints and safety endpoints. Animal data is often reported for mechanisms and “unconditioned anxiety.” A systematic meta-analysis of SSRIs in animal models supports efficacy in the unconditioned anxiety model (Heesbeen et al., 2024, PMID 38980348). But that is not the same as a reliable prediction for humans: animal models only capture certain aspects, and translatability can be clearly limited.

Observational and cross-sectional data in anxiety are especially prone to confounding (e.g., lifestyle, comorbidities, socioeconomic factors). RCTs reduce that bias more effectively. So, if you want to estimate “works,” you should prioritize RCT-based evidence whenever possible and treat observational data more as contextual information.

What does this mean practically for you?

If an intervention performs well in RCTs and is consistent across multiple studies in a meta-analysis, it provides a robust foundation. If you only have animal data or only observational data, the evidence base is much weaker. And if a meta-analysis finds effects but includes highly heterogeneous populations or different measurement instruments, interpretation should be conservative.

Lifestyle first: Which interventions change anxiety symptoms and closely related targets

Short answer: For anxiety, non-drug approaches are especially plausible in several meta-analyses because they often target both symptoms and close endpoints like sleep. Tai Chi has shown effects in older adults, but the exact “dose range” depends on the study evidence base.

If you want to prioritize anxiety-related effects, a practical starting point is: lifestyle levers first—not because it’s a trend, but because in studies they often address multiple outcomes simultaneously. That also reduces the chance that an effect is based on a single random measurement point.

Tai Chi in older adults (Bayesian meta-analysis)

Chen et al. report in a Bayesian meta-analysis of Tai Chi for anxiety, depression, and sleep quality in older adults (Chen et al., 2025, PMID 40625025). Bayesian approaches are particularly relevant because they model uncertainty and estimate “optimal dosing ranges” statistically. Still, even if an optimal dosing range is identified, it remains dependent on the included study design (e.g., training frequency, duration, and follow-up). For your self-decision: use the study as a signal for effectiveness and as a starting point for parameter selection, but check in the original work what training structure was actually tested.

Anxiety in dementia: combined effectiveness is possible, but population-dependent

For people with dementia in everyday life, systematic reviews and meta-analyses suggest that both pharmacological and non-pharmacological approaches can be clinically effective (Nimmons et al., 2024, PMID 38097097). At the same time, effect size and transferability depend on the subpopulation studied and the study selection criteria. This matters because anxiety in dementia often arises in a different context (cognitive burden, environment, daily structure). An “anxiety therapy” there can simultaneously address an environmental and communication problem—something to consider in your planning.

Young adults and “eco-emotions” (causal question remains open)

Rana et al. study the relationship between so-called eco-emotions and mental wellbeing in young adults (Rana et al., 2026, PMID 42089230). Even if this fits thematically into the anxiety/stress landscape, the central question remains: is it causal, or do the effects run through third variables (e.g., media exposure, lifestyle, social environment)? This causal gap is common in complex psychological topics.

Why movement/structured practice is a sensible lever

Movement and physically structured practice are often evaluated not only for anxiety symptoms, but also for sleep and mood—and that increases the chance that effects are multidimensional. This doesn’t mean movement works equally strongly against every type of anxiety; rather, the evidence base is often broad enough that you can improve close targets (e.g., sleep quality), which may then indirectly influence anxiety.

Age and context differences

Important: What works in older adults doesn’t necessarily work 1:1 in other age groups. Even in dementia, baseline conditions differ. The evidence base is encouraging, but it isn’t “universal.” Therefore, your decision should be tied to your starting point: disorder type, comorbidities, sleep patterns, and your current life reality.

Medications and related therapies: What meta-analyses for anxiety really say

Short answer: Medication data can’t be assessed as simply “effective” or “ineffective” because endpoints and safety reporting differ. Animal data on SSRIs support effects in unconditioned anxiety, but they don’t guarantee similar effects in humans. Meta-analyses can often provide a ranking within specific populations more than a generalizable statement.

Medications are often used for anxiety, but the scientific interpretation requires precision: which anxiety type, which outcome, and which safety profile? In practice, anxiety trials are often more complex than “one pill, one effect.”

SSRIs: Evidence from animal models (unconditioned anxiety)

Heesbeen et al. systematically summarize animal data on how SSRIs influence unconditioned anxiety (Heesbeen et al., 2024, PMID 38980348). This is an important preliminary step because it provides biological plausibility and model consistency. But: an animal model doesn’t capture the full complexity of human anxiety (learning processes, cognitive appraisals, real-life situations, comorbidities). Therefore, from this meta-analysis you can mainly conclude: “There are consistent effects in the animal model.” Not reliably: “In humans, the effect is the same size or the same speed.”

Network meta-analysis: ranking instead of general truth

Tan et al. conduct a network meta-analysis of interventions for anxiety and depression in a specific population (PCOS) (Tan et al., 2026, PMID 41664652). Network meta-analyses are useful for comparing interventions relative to one another when direct head-to-head comparisons are missing. But the strength of the conclusions depends on:

• study quality and comparability,
• heterogeneous measurement instruments,
• and whether the interventions were truly primarily designed to target anxiety.

That means the study isn’t “useless”—it just isn’t automatically transferable to every anxiety diagnosis.

Anxiety-specific endpoints vs. secondary effects

A common reason medication trials can produce mixed results: some studies primarily treat other symptoms (e.g., sleep, mood, daily structure, or aspects of comorbid illness). If anxiety is measured only as a secondary outcome, effects may be smaller, mixed, or more heterogeneous. That means a meta-analysis can still combine effects statistically, but you should be more skeptical when anxiety wasn’t a primary target.

Safety assessment: prefer RCT-based reviews

Safety is especially relevant in anxiety (e.g., tolerability, adverse event profiles, interactions). The right strategy is: assess safety questions as much as possible via reviews that systematically extract RCTs and adverse-event data—not via indirect inferences from other endpoints. The principle of “extracting side effects cleanly from RCTs” is demonstrated in systematic work focusing on safety in RCT appraisal (Jalal et al., 2026, PMID 41324388). Translated to anxiety: if adverse events are not consistently reported in the RCTs, your safety estimation is limited.

Practical implication

If you decide on medication, the choice should be tied to your specific diagnosis and your target size: symptoms, functional ability, or e.g., sleep. It should also involve medical supervision, because safety and interaction questions are individual. Meta-analyses can help you understand the probability of benefit and the remaining uncertainty, but they don’t replace a personalized risk–benefit assessment.

Nutrition, GLP‑1, and special topics: Where the data is interesting—and where it remains limited

Short answer: For ketogenic diets, GLP‑1 receptor agonists, and other special topics, there are interesting hints, but the data on anxiety overall is limited or context-dependent. A “side-effect/risk signal” is not the same as a well-established treatment effect against anxiety.

Ketogenic diets: systematic evidence on depression and anxiety

Janssen-Aguilar et al. summarize in a systematic review and meta-analysis ketogenic diets in the context of depression and anxiety (Janssen-Aguilar et al., 2026, PMID 41191382). That sounds direct at first—but the key issue is how well the included studies actually measured anxiety endpoints. In meta-analyses, the strength of the conclusion can strongly depend on whether anxiety was truly examined as a primary outcome or whether it was mostly related endpoints. If anxiety wasn’t operationalized cleanly in the original studies, the evidence for anxiety specifically remains limited.

For you: even if a ketogenic diet is emotionally/mood-useful for some people, the data for anxiety as a target (with clear dose–time relationships) may not be robust enough to treat it as an “anxiety therapy.”

GLP‑1 receptor agonists: risk for psychiatric disorders—not “anxiety treatment”

Han et al. report in a systematic review and meta-analysis of RCTs the association between GLP‑1 receptor agonists and the risk of psychiatric disorders (Han et al., 2026, PMID 41852258). This matters because it provides a safety/risk perspective. But it doesn’t automatically mean GLP‑1 agonists either treat anxiety effectively or cause anxiety “in the sense of a direct anxiety therapy.”

Here, interpretation is also tricky: “risk” in pharmacovigilance/study analyses can mean that certain events were observed more frequently in the RCTs—or that differences existed in how outcomes were recorded. Without a clear, anxiety-focused efficacy test, this remains a safety/health risk signal—not evidence of a therapeutic effect against anxiety.

Eco-emotions: association instead of clear causality

Rana et al. examine eco-emotions and mental health in young adults (Rana et al., 2026, PMID 42089230). Even if the topic addresses emotionally distressing themes, the causality question is often not settled: association isn’t the same as cause. As a result, for concrete self-decisions (e.g., “Should I do X to reduce anxiety?”), it currently can’t serve as a direct treatment instruction.

Contextualization: Why special topics shouldn’t replace therapy decisions

Special topics can provide signals—but they shouldn’t replace anxiety-focused treatment decisions based on direct endpoints. If you want to treat anxiety as the main problem, interventions with clear anxiety outcomes (e.g., in studies of psychological approaches or physically structured programs) are usually the better foundation.

Weighing safety and potential benefit

If you consider such topics at all (e.g., very strict nutrition or GLP‑1 medications), you should separate the evidence:

• Is it a treatment effect for anxiety? (often limited)
• Or is it a safety/risk issue? (then interpret differently)
• How was anxiety measured? (crucial for the validity of conclusions)

Study overview with evidence interpretation: What you can infer (and what you can’t)

Short answer: The available meta-analyses provide mostly context-dependent indications: for lifestyle programs, there are often more consistent signals on anxiety-adjacent endpoints, while special topics regarding anxiety are frequently limited or indirect. Animal data supports mechanisms, but transfer to humans is limited.

How to read the table correctly

• Animal data: Even if an effect is robust in an animal model, transferability remains limited. This isn’t a “weakness” of animal research—it’s a methodological boundary. (Heesbeen et al., 2024, PMID 38980348)
• Population-specific analyses: In PCOS or dementia, results are often not fully transferable 1:1 to other groups. (Tan et al., 2026, PMID 41664652; Nimmons et al., 2024, PMID 38097097)
• Nutrition/special topics: Association doesn’t automatically mean a treatment effect for anxiety. Ketogenic diets can be interesting, but the conclusion for anxiety depends on how well anxiety was operationalized in the primary studies. (Janssen-Aguilar et al., 2026, PMID 41191382)
• Risk rather than effect: For GLP‑1 agonists, the meta-analysis focuses on psychiatric disorders as a safety/risk association—this must be separated from “anxiety is treated.” (Han et al., 2026, PMID 41852258)

What you can infer (and what you can’t)

You can infer:

• Which interventions in specific settings show more consistent improvements.
• Where knowledge gaps exist (e.g., anxiety endpoints are rarely studied as primary outcomes).
• Where safety questions should be prominent.

You shouldn’t infer:

• A universal “ranking list for anxiety” without context.
• Direct transfer of animal or special-topic data to your personal anxiety diagnosis.
• A cause-and-effect judgment from associations when causality wasn’t tested.

If you want to go deeper, use this overview as a starting point and check the original papers for which endpoints were measured (anxiety vs. sleep vs. mood) and how much uncertainty remains.

What you should take away

• Anxiety is not an outcome. Study design, population, and measurement instruments determine what “effectiveness” in a meta-analysis actually means.
• Lifestyle often has the closest practical evidence, because interventions like Tai Chi and structured physical practice frequently involve multiple close targets (e.g., sleep) as well (Chen et al., 2025, PMID 40625025).
• Animal data on SSRIs supports mechanisms in unconditioned anxiety, but it isn’t a direct prediction for humans (Heesbeen et al., 2024, PMID 38980348).
• Special topics are limited or context-dependent: ketogenic diets and GLP‑1 receptor agonists provide more hints/risk signals, not a robust basis for an anxiety-treatment decision based on direct anxiety endpoints (Janssen-Aguilar et al., 2026, PMID 41191382; Han et al., 2026, PMID 41852258).
• Use meta-analyses as a map, not as the goal. For an informed decision, you must always verify which endpoints and target groups were actually studied.