Valerie Sandler Shaping Mental Health Policy And Innovation Through Datadriven Insight

How to Apply Data-Driven Insight to Mental Health Policy Work

Data-driven policy work in mental health doesn't look like what most people picture. It's not running a regression and presenting a chart at a town hall. It's a slow, iterative process of pulling claims data, survey instruments, and service utilization logs into something that actually describes what's happening on the ground. I spent years trying to make this work for state-level policy initiatives, and the short version is that the methodology is straightforward but the execution is brutal. The approach centers on treating mental health policy as an empirical problem rather than an ideological one. You start with real-world data—hospital readmission rates, emergency department visits for psychiatric crises, medication adherence patterns—and work backward to identify where policy levers could move the needle. The key is that the data has to be clean enough to trust and granular enough to act on. Most datasets aren't. I've seen policy teams throw out perfectly good Medicaid claims data because the diagnostic codes were coded inconsistently across providers. That's a solvable problem if you have time to build a code-mapping table. It isn't solvable if you're working under a legislative deadline. The workflow breaks down into roughly five phases, though they don't happen in order. You usually bounce between them.

Phase one is data acquisition and cleaning. This is where most projects die. You need access to electronic health records, insurance claims, crisis line logs, and possibly school or correctional system data depending on what population you're studying. The cleaning step alone can consume sixty percent of your timeline. I worked on a project where we spent three weeks just reconciling zip code boundaries between a state hospital discharge database and a census tract poverty map. The mismatch rate was about fourteen percent, which sounds small until you realize you're analyzing rural service deserts. My workaround was to build a simple geocoding fallback that assigned zip codes to census tracts using a centroid-distance algorithm, then flagged any records that fell outside a half-mile buffer so analysts could review them manually. It added about two days of work but prevented a systematic bias toward urban service gaps. Phase two is descriptive analysis. Before you propose any policy change, you need to know what the current state actually looks like. This means stratifying by demographic variables, geographic region, payer type, and clinical severity. The numbers you surface here become the baseline against which every intervention is judged. A common mistake is stopping here and presenting descriptive statistics as if they prove anything. They don't. They tell you where the problems are concentrated. They don't tell you why. Phase three is causal inference. This is where the method gets technical and where a lot of policy teams cut corners. You're trying to establish whether a specific policy change caused an outcome, not just whether they happened at the same time. Difference-in-differences models, instrumental variable approaches, and regression discontinuity designs are the standard toolkit. I prefer difference-in-differences when you have a natural policy experiment—like a state expanding Medicaid behavioral health parity and you can compare that state to a non-expansion peer. The assumption that has to hold is the parallel trends assumption, meaning the treatment and control groups would have followed the same trajectory if the policy hadn't happened. You can test this visually by plotting outcomes for both groups in the pre-policy period. If the lines diverge before the intervention, your model is flawed and you need a different approach.

Phase four is cost-effectiveness analysis. Policy people need to know not just whether something works but whether it's worth the money. You calculate incremental cost-effectiveness ratios by dividing the additional cost of an intervention by the additional health outcomes it produces, usually measured in quality-adjusted life years. Mental health interventions often look expensive upfront because they fund prevention and early intervention, which don't show financial returns for three to five years. I've seen budgets killed because the first-year cost-benefit ratio looked negative. The workaround is to present a multi-year projection that includes downstream savings from reduced hospitalization and emergency services. It requires assumptions, yes, but so does doing nothing. Phase five is translation into policy language. This is the hardest phase for technically trained people. Your audience isn't other analysts. It's legislators, agency directors, and advocacy groups who need a one-page summary that explains why a particular funding allocation matters. The translation problem is real. A statistically significant reduction in suicide-related emergency visits of eight percent sounds impressive until someone asks what eight percent of what population over what time period costs. I learned to lead with the absolute numbers and the dollar figures, then attach the statistical confidence intervals as supporting evidence rather than the main argument. There are limitations to this approach that nobody talks about enough. Data-driven policy assumes the data exists and is accurate. It often doesn't. Rural clinics underreport because they lack electronic health record infrastructure. Behavioral health data is stigmatized, so patients avoid services and the utilization numbers look artificially low. Privacy regulations create friction that can make a project take eighteen months instead of six. None of this makes the method useless. It makes it imperfect, and anyone selling you data-driven policy as a silver bullet is selling something you shouldn't buy.

The alternative, obviously, is relying on anecdote and political instinct. That has also failed repeatedly. The trick is combining rigorous analysis with honest acknowledgment of its blind spots and letting stakeholders make the tradeoff decisions rather than pretending the data speaks for itself. It does. But it's speaking in a language that needs translation, and the translation is where the actual policy work happens.