Protocol Generator · TIT
Trend-in-Trend protocol
In plain terms. A drug's use rises steeply in some kinds of patient and barely at all in others. If the drug really causes the outcome, the outcome should rise fastest exactly where the drug's use rose fastest. Trend-in-trend reads the effect off that co-movement rather than off a comparison of treated with untreated people — so a confounder that sits still over time cannot produce it.
How that is done precisely (Ji et al, Epidemiology 2017)
Stage 1 fits a logistic model of exposure on covariates that affect exposure but are not affected by it, and takes each person's predicted probability of being exposed at any point in the study period — the cumulative probability of exposure (CPE). Covariates are treated as time-invariant, so each person sits in one CPE stratum for the whole study. The population is then cut into quantiles of CPE (quintiles in the original paper), which sorts people by how steeply their exposure prevalence trends. Stage 2 tabulates, for every CPE stratum and calendar period, a 2 × 2 table of exposed / unexposed × case / non-case, and maximises the trend-in-trend likelihood over those tables. The estimate is exp(β₁), an odds ratio.
When it fails. The design is unbiased unless there are unmeasured factors affecting the outcome whose prevalence trends over calendar time in a way that is correlated with the exposure trend across CPE strata. A confounder that is stable over time therefore cannot bias it; a co-intervention or a secular change that arrives alongside uptake can. It also needs a genuinely strong time trend in exposure — without one there is nothing to read the effect off, and the likelihood may not be identified.
Live protocol preview
Consistency checks · TIT 防呆
Study design diagram
Phenotype library
Pick a phenotype
Each entry says whether a study has measured how well it identifies what it names. Nothing here is added to your protocol's reference list — this builder has no mechanism for that, so cite what you use yourself.