Is Mental Disease Overdiagnosed or Better Recognized?

Isolated prompt-specific note. This applies the diagnosis-as-attribution framework directly to the question. It is kept separate from the four general memos because its structure is an answer to the JLC prompt rather than a standalone piece of the VMS framework.

More diagnoses do not by themselves demonstrate either overdiagnosis or better recognition.

diagnostic expansion = a descriptive change in reach or rate
diagnostic quality = the welfare value of the attributions made

An increase in diagnoses can come from a change in the prevalence of relevant states, help-seeking, access, screening, disclosure, diagnostic thresholds, institutional contact, professional practice, categories, or the probability of receiving a specific attribution after contact. None of these changes settles whether the additional diagnoses should have been made.

The prompt is therefore better restated:

Is the diagnostic system becoming better or worse
at selecting the attributions that should be made?

This is a question about diagnostic selection efficiency, not raw prevalence.

Why the number expands.

For a specific attribution $a$, the observed diagnosis rate can be decomposed as:

$$P_t(A=a) = \sum_{s\in S_t} P_t(A=a\mid H=1,s) P_t(H=1\mid s) P_t(s),$$

where:

$$P_t(s)=\text{prevalence of baseline condition-state }s,$$ $$P_t(H=1\mid s)=\text{probability of seeking or reaching help given }s,$$

and:

$$P_t(A=a\mid H=1,s)=\text{probability of receiving attribution }a\text{ after contact}.$$

The total diagnosis rate is:

$$P_t(A\neq\varnothing) = \sum_{a\neq\varnothing}P_t(A=a).$$

Expansion may therefore reflect:

$$\Delta P(s),$$ $$\Delta P(H=1\mid s),$$

or:

$$\Delta P(A=a\mid H=1,s).$$

Help-seeking is not merely a supply-side variable. Patients may reinterpret their experience, disclose symptoms, pursue documentation, seek care, or use diagnostic language autonomously. For a historical comparison centered on changes in the medical-diagnostic system, it may still be analytically useful to treat autonomous demand-side change as secondary and focus on categories, thresholds, professional practice, access pathways, institutional gatekeeping, and welfare-conversion capacity. That is a simplifying assumption, not a denial of patient agency.

Reach is diagnosis-specific.

Different diagnoses can expand into different parts of the population. For attribution $a$, define reach as:

$$R_t(a) = \{s\in S_t:P_t(A=a\mid H=1,s)P_t(H=1\mid s)>0\}.$$

Total diagnostic reach is:

$$R_t=\bigcup_{a\neq\varnothing}R_t(a),$$

and expansion is:

$$\Delta R=R_1-R_0.$$

An expansion in ADHD, autism, depression, anxiety, psychosis, trauma-related diagnosis, or personality diagnosis may have a different welfare composition. There is no reason to assume one common trend merely because the total number rises.

Conversion capacity can enlarge the target set.

Better treatment, accommodation, protection, support, prognosis, self-understanding, or blame reduction can make recognition valuable for states that were previously not worth formally diagnosing.

conversion capacity rises
→ more presentations can benefit from attribution
→ the welfare-positive diagnostic target set expands

This is a real form of better recognition. A mild, unusual, or previously ignored state can become worth recognizing once diagnosis opens a useful pathway.

But expansion often reaches more marginal and context-dependent presentations. These may be harder to distinguish from ordinary distress, grief, developmental variation, self-expression, environmental mismatch, or social failure.

target set expands
→ boundary selection becomes harder
→ better recognition and overdiagnostic pressure can increase together

The two phenomena are not mutually exclusive descriptions of one count. The same period can correct historical under-recognition in some regions while creating new boundary errors in others.

Two failure modes.

Boundary overdiagnosis occurs when a state–source relation should not have entered diagnosis in that way. No formal attribution, a different social instrument, or a non-diagnostic intervention would have produced better welfare. The system assigns a problem to non-self pathology when its important provenance lies in ordinary variation, adverse environment, social conflict, self-expression, or another domain.

Attributional failure occurs when some form of diagnosis could be welfare-positive, but the actual attribution is inaccurate, too coarse, too precise, or directed at the wrong source, level, or pathway. The appropriate response may be better diagnosis rather than less diagnosis.

wrong entry into diagnosis → boundary problem
wrong attribution after justified entry → attribution problem

The historical classification of homosexuality illustrates the distinction. Distress related to sexuality could be real and welfare-relevant. That did not make homosexuality itself the pathological source. The attribution misplaced harm generated by stigma, pressure, identity conflict, and hostile environment inside the person or identity being classified. Real distress did not validate the source attribution.

Individual diagnostic value.

For an attribution $a$ and presentation $x$, the relevant comparison is:

$$\delta W^{soc}(a,x;\Omega_t^{conv}) = W^{soc}(o(a,x;\Omega_t^{conv})) - W^{soc}(o(\varnothing,x;\Omega_t^{conv})).$$

If $\delta W^{soc}>0$, the attribution is welfare-positive relative to no formal attribution. If $\delta W^{soc}<0$, it is welfare-negative.

For refinement among already diagnosed cases, the comparator should be the earlier or alternative attribution rather than $\varnothing$. A reach-expansion problem and a refinement problem therefore require different counterfactuals.

A standard narrow understanding of overdiagnosis focuses on diagnoses whose harms and costs exceed their benefits. That maps onto the strongly welfare-negative case: removing the diagnosis without replacement raises welfare.

A beneficial diagnosis may still be non-optimal under scarcity. That is diagnostic inefficiency or misallocation rather than necessarily a strongly harmful diagnosis. The distinction matters because a system can fail by making harmful attributions, by choosing lower-value positive attributions over higher-value ones, or by missing valuable candidates.

Bundles, interaction, and feasible selection.

At a time-slice $t$, let $D_t$ be the set of possible diagnostic decisions, $R_t^{res}$ the relevant resource constraints, and $B$ a feasible diagnostic bundle. Because diagnoses may be substitutes or complements, the system should evaluate the bundle as a whole:

$$W(B),$$

without assuming:

$$W(B)=\sum_{d\in B}W(d).$$

The optimal feasible bundle is:

$$B_t^*=\arg\max_B E[W(B)] \quad\text{subject to }R_t^{res}.$$

Let $B_t$ be the bundle selected by the actual framework. A cross-sectional efficiency measure is:

$$\mathrm{Eff}_t = \frac{E[W(B_t)]}{E[W(B_t^*)]}.$$

This represents the actual expected welfare generated by the system relative to the true maximum feasible expected welfare under the same constraints.

Opportunity cost is already present in the feasible-bundle comparison. After comparing $W(A)$ and $W(B)$, the forgone bundle should not be subtracted again. Doing so would count the same opportunity cost twice.

The benchmark is the true frontier, not current institutional belief.

The denominator should not be “the best bundle according to the institution's present evidence.” The purpose of the benchmark is to evaluate whether the diagnostic framework actually identifies the cases and attributions it should identify.

If epistemic failure is built into the optimum, the benchmark loses its critical force.

best available evidence may explain failure
best available evidence does not define the true optimum

Diagnostic efficiency therefore includes epistemic accuracy. A framework's inability to identify welfare-positive diagnostic opportunities is part of its inefficiency, even when the failure is understandable.

The benchmark uses true expected welfare, not fully observed lifetime welfare. Following every downstream consequence would be impossible. At each time-slice, each feasible bundle can nevertheless be understood as having a true probability-weighted welfare value, whether or not institutions can estimate it perfectly.

Attributional frontier.

The bundle formulation can also be expressed at the level of attribution maps. Let:

$$g_t^{act}:X_t\to A\cup\{\varnothing\}$$

be the actual system, with yield:

$$Y_t^{act} = \int_{X_t} \delta W^{soc}(g_t^{act}(x),x;\Omega_t^{conv}) \,dF_t(x).$$

Let $\mathcal G$ be the general, non-time-indexed space of possible attribution systems. It is not limited to current categories, thresholds, policy classifications, or professional practice. The theoretical frontier is:

$$Y_t^* = \sup_{g\in\mathcal G} \int_{X_t} \delta W^{soc}(g(x),x;\Omega_t^{conv}) \,dF_t(x).$$

Attributional efficiency is:

$$E_t=\frac{Y_t^{act}}{Y_t^*}, \qquad Y_t^*>0.$$

Holding $\mathcal G$ fixed, the frontier can still change across time because the presentation distribution $F_t(x)$ and the conversion environment $\Omega_t^{conv}$ change. New treatments, accommodations, resource conditions, institutions, and social responses can change which attributions are worth making and how much welfare they can generate.

Why compare efficiency rather than raw welfare across time.

Raw welfare generated by diagnosis is not directly comparable across periods because many background conditions change:

resources and clinical capacity
treatment and drug availability
laws, insurance, and school systems
social stigma and public awareness
economic and institutional conditions

The better comparison is the diagnostic system's performance relative to its own time-slice frontier.

Diachronic change is:

$$\Delta E = E_1-E_0 = \frac{Y_1^{act}}{Y_1^*} - \frac{Y_0^{act}}{Y_0^*}.$$

If efficiency rises, the system moves in a better-recognition direction:

$$E_1>E_0.$$

If efficiency falls, the system moves in an overdiagnostic or diagnostically inefficient direction:

$$E_1<E_0.$$

A fall may involve more strongly welfare-negative diagnoses, more non-optimal diagnoses under scarcity, worsening attribution, resource misallocation, or failure to recognize higher-value candidates.

The marginal value of expanded reach.

Let the newly reached set be:

$$M_1=R_1-R_0.$$

For pure reach expansion, the default comparison is no formal attribution:

$$\Delta Y^{exp} = \int_{M_1} \delta W^{soc}(g_1^{act}(x),x;\Omega_1^{conv}) \,dF_1(x).$$

Equivalently:

$$\Delta Y^{exp} = \int_{R_1-R_0} \left[ W^{soc}(o(g_1^{act}(x),x;\Omega_1^{conv})) - W^{soc}(o(\varnothing,x;\Omega_1^{conv})) \right] \,dF_1(x).$$

If $\Delta Y^{exp}>0$, the expansion is welfare-positive recognition. If $\Delta Y^{exp}<0$, the expansion is welfare-negative or overdiagnostic.

The marginal value can be decomposed into:

$$\Delta Y^{exp} = \Delta W^{coarse+} + \Delta W^{refined+} - \Delta C^{coarse-} - \Delta C^{misattribution} - \Delta C^{dilution/resource}.$$

These terms represent useful coarse recognition, useful refined attribution, low-value coarse labeling, attribution to the wrong source or level, and dilution or resource crowding.

Better recognition dominates when:

$$\Delta W^{coarse+}+\Delta W^{refined+} > \Delta C^{coarse-}+\Delta C^{misattribution}+\Delta C^{dilution/resource}.$$

Overdiagnostic harm dominates when the inequality reverses.

The resulting answer is necessarily mixed and category-dependent. Diagnostic expansion can contain genuine recognition, direct harm, imprecise but improvable attribution, and scarcity-based misallocation at the same time. The original binary is misleading because it asks one count to choose between several distinct changes in diagnostic quality.

notes on text Brodersen et al. provide a working definition of overdiagnosis and distinguish it from false positives, overtreatment, overtesting, and misdiagnosis. Drescher provides the historical material on homosexuality and DSM depathologization. Garrett's discussion of attribution and psychiatric response is close to the intervention question considered here.