Design: analysis of two prospective multicentre cohorts with mixed-effects ordinal regression and Cox models
Population: 25,602 patients in NS-Park and 1,441 in PPMI with Parkinson's disease
Primary outcome: freezing of gait severity and incidence in relation to levodopa exposure
Effect: adjusted odds ratio for levodopa 1.84 (95% CI 0.94-3.61) in NS-Park and 0.76 (0.38-1.51) in PPMI; disease duration 4.4 (3.01-6.29) and 6.97 (5.05-9.63)

Two prospective cohorts were analysed: 25,602 patients with Parkinson's disease in NS-Park and 1,441 in PPMI, with freezing of gait scored on MDS-UPDRS Part II item 2.13 and nested subcohorts defined by disease duration, baseline freezing and levodopa exposure.

Unadjusted, the familiar association appeared in NS-Park - more levodopa, more freezing - and was absent in PPMI. After adjustment for disease duration, Hoehn and Yahr stage and MDS-UPDRS Part III, levodopa was not significantly associated with freezing severity in either cohort, and the two point estimates pointed in opposite directions: odds ratio 1.84 (95% CI 0.94 to 3.61) in NS-Park and 0.76 (0.38 to 1.51) in PPMI. Disease duration dominated, at 4.4 (3.01 to 6.29) and 6.97 (5.05 to 9.63), as did motor severity. Future levodopa exposure did not predict de novo freezing; incident freezing was driven by Hoehn and Yahr stage and motor score.

The clinical consequence is about counselling rather than prescribing. Levodopa-sparing strategies adopted specifically to avoid freezing have no support here, and the fear of levodopa that leads patients and some clinicians to delay adequate dosing costs mobility that could have been had. Freezing tracks the disease, not the drug.

Do not delay or restrict levodopa in order to prevent freezing of gait
Say plainly that freezing reflects disease stage and duration, which is what patients most often ask about
Record Hoehn and Yahr stage and motor score - those are what predict freezing
Distinguish off-period freezing, which responds to dose optimisation, from on-period freezing, which does not
Refer for gait-directed physiotherapy and cueing strategies rather than reaching for a drug change

Why it matters

It removes a reason clinicians and patients have used to under-treat Parkinson's disease.

The statistics, in plain English

The important comparison is between the unadjusted and adjusted results. Unadjusted, levodopa looks harmful because patients with longer, more severe disease take more of it and also freeze more - textbook confounding by indication. After adjustment the confidence intervals cross 1.0 in both cohorts and point opposite ways, which is what a null result looks like when it is genuinely null rather than merely underpowered. Note this is still observational: adjustment reduces confounding, it does not remove it, and residual confounding by severity is the standing caveat.