Design Experiments From Reported Results

 Long Analysis Plan Copied

Evidence base for the sibling-comparison design

The supplied evidence records do not directly address the SSRI–miscarriage question in pregnancy; they cover SSRI sexual dysfunction, apathy, self-concept changes, microglial anti-inflammatory effects, and SSRI–TRKB pharmacokinetics. No study in the record quantifies miscarriage risk by SSRI exposure or by sibling-discordant exposure. What the records do support is why the confounding/indication concern is serious: depression itself (the indication) and treatment patterns vary within families and within mothers over time, so an observed SSRI–miscarriage association can reflect shared familial liability, shared psychiatric severity, or unmeasured lifestyle/environmental factors rather than a direct drug effect.

A defensible within-family design in the Danish registries (Danish National Patient Registry, Danish Medical Birth Register, Danish Psychiatric Central Register, Statistics Denmark family links, and the Danish National Prescription Registry for SSRI fills) would proceed as follows.

Recommended analysis plan

• Cohort assembly: Identify all singleton pregnancies 1997–2023 (SSRI era), linked to mothers via Statistics Denmark family IDs.
• Exposure window: SSRI dispensing from 30 days before conception through end of gestation; secondary definitions restricting to fill during weeks 4–12 (placental transfer window) versus later to distinguish early-pregnancy drug effect from indication-driven confounding.
• Sibling strata: Restrict to mothers with ≥2 recorded pregnancies; classify each pregnancy as SSRI-exposed or unexposed using the strict window. Both concordant (both exposed / both unexposed) and discordant (one exposed, one unexposed) pregnancies contribute to the marginal (population) estimate; only discordant strata inform the within-family estimate.
• Primary model: Conditional logistic regression (cluster on mother/family) for miscarriage (<24 weeks) with covariates: maternal age at each pregnancy, psychiatric diagnosis at conception, co-medication (other psychotropics), smoking, parity, pregnancy spacing, calendar year. Within-family (fixed-effects) model drops all time-invariant familial confounders (shared genetics, household environment, parental socioeconomic status).
• Indication-adjustment layer: Add stratum-level indication severity proxies (inpatient vs outpatient diagnosis, prior psychiatric admissions) to test how much residual association tracks psychiatric severity rather than drug exposure. If the within-family estimate collapses toward the null once indication is modeled, the observed association is plausibly confounded by indication.
• Quantifying attenuation: Report the ratio of the within-family conditional OR to the population (unmatched) OR; attenuation = 1 − (within OR / population OR). A large attenuation (e.g., 60–80% drop) indicates familial/genetic/environmental confounding dominates; minimal attenuation with precise CIs is more consistent with a direct drug effect. 0.7 (example: a within-family OR of 0.30 relative to population OR of 1.0 would represent 70% attenuation — this is illustrative, not observed data).
• Sensitivity analyses: (a) Include both exposed/unexposed in each stratum via alternate assignment; (b) stratify by psychiatric diagnosis timing (pre-existing vs de novo during pregnancy); (c) test survival bias by restricting to live-birth siblings only, which biases against detecting miscarriage in exposed pregnancies and must be explicitly modeled via Heckman-style correction or inverse-probability weighting on pregnancy inclusion.

Key statistical risks specific to this question

Survivor bias is the dominant threat: pregnancies ending in miscarriage are less likely to produce subsequent siblings in a discordant comparison that only uses recorded births. Danish pregnancy-loss records (via obstetric discharge diagnoses and, historically, incomplete early-loss capture) under-ascertain losses before 12 weeks, which inflates apparent miscarriage-free survival in both exposure groups. Sample size is the second major constraint: the number of mothers with ≥2 pregnancies and discordant SSRI exposure is a small fraction of the full SSRI-exposed cohort; precision on the within-family OR will be wide unless ≥several thousand discordant strata are available. Power analysis must precede any interpretation.

What the supplied evidence does and does not support

The records confirm psychiatric-confounding plausibility (e.g., within-family variation in treatment decisions and self-concept around SSRI use Danish qualitative study of SSRI treatment decisions) and anti-inflammatory properties of SSRIs that could plausibly influence pregnancy biology SSRI anti-inflammatory microglial effects. Neither record reports miscarriage outcomes; direct effect-size data for the SSRI–miscarriage association are not present in the supplied evidence and must come from the registry analysis itself.

What would falsify or revise the planned interpretation

• If the within-family OR remains ≈1 after adequate power (n_discordant_strata sufficient), familial/indication confounding fully explains the observed association.
• If the within-family OR is attenuated but non-null, a residual direct drug effect remains plausible; the key residual confounders are psychiatric severity fluctuations within the same mother (since indication is NOT time-invariant) and smoking/medication changes between pregnancies.
• If the population estimate is null but the within-family estimate is elevated, this would signal exposure-carryover bias (e.g., pregnancies conceived during tapering classified as unexposed) and would require re-specification of the exposure window.

Note: The bar chart above is a planning illustration of the expected attenuation pattern — not actual observed data. No sourced numbers for SSRI–miscarriage ORs are present in the supplied records; actual values must come from the registry analysis.