The data we deploy is constructed longitudinally, at the firm level using the establishment-level, annual database on employment and sales from National Establishment Time Series for the years 1990 through 2015. These form the basis for matching recipients of incentive offers to firm-level data and identification strategies.
We obtained data on incentives from the Michigan Economic Development Corporation, on nine types of incentive programs. These are the 21st Century Jobs Fund, Community Development Block Grants, Michigan Business Development Program, Michigan Economic Growth Authority, other Michigan Business Tax credit approvals, private activity bonds, renaissance zones, seed capital program funds and other uncategorized grants and loans.
The process required significant data matching and reconciliation. For example, our initial sample included 4,217 incentives across those nine programs. There were missing approval dates for 215 listed incentives. We excluded these firms from both the sample and control groups.
We manually identified the unique Dun & Bradstreet identifier, or DUNS numbers, for each establishment that was offered an incentive in Michigan from the NETS database. Out of 4,002 incentive deals, we were able to match the DUNS numbers for 2,997 of them. Among these, 1,962 single incentives were offered over the study period; while the other 1,035 firms were offered multiple awards across categories over time. After removing unmatched DUNS numbers, our final study sample had 1,890 firms that were offered only one incentive during the study period and 412 additional firms that received multiple incentive offers over time.
To eliminate any bias that might appear in the results from including firms that received multiple offers, we started the analysis with those firms that received only one incentive offer (N=1,890 establishments). Further as a robustness test, we did an additional analysis that included multiple incentive firms to our analysis (N=2,302 establishments) and estimate the effects.
We attempt to estimate the impact of incentives on employment and sales data. In order to account for endogeneity, we created a representative control group to isolate the treatment group’s impact from contemporaneous changes in establishments. We selected firms for the control group using a propensity score matching without replacement method. In this way we identified a total of five controls for each treated establishment by also including matched observed covariates: type of industry (relying on firms’ 2-digit SIC code), establishment category (whether the firm is a branch, headquarters or standalone), subsidiary firm or not, and establishment size (very small, small, medium, or large firms). The caliper width was set at 25% of standard deviation of the propensity scores.
We then convert the NETS data to a panel dataset to exploit the variation over time. We follow Donegan, Lester and Lowe to identify our models. We create our variable of interest — Incentive — which is equal to 1 if the current year is greater than or equal to the award approved year for the treated group. This variable is equal to zero if the establishment is a control group or the current year is less than the award year for the treated group.
With an identification strategy that includes both treatment and control group, our model specification is a straightforward treatment test, taking the form:
Υit ≡ α + βIncentiveit + ylinear time trendt + δi + λt + Matching groupt + εit (1)
Where subscripts i and t represents establishments and years, respectively. Υ is the outcome variable of interest such as establishment employment and sales. We express the sales in 2015 dollars. The coefficient of interest, β, estimates the causal impact of offered incentives on establishment outcomes.
We include linear time trends to control for any unobserved trends that are common for all establishments. We include establishment fixed effects (δi) to control for heterogeneity across establishments. In alternate specifications, we also relax this assumption and estimate random effects. We also include year fixed effects (λt) to control for factors that cause year-to-year changes across all establishments. Further, we also control for matching group-specific time trends to capture unobserved factors that vary over time for a treated and its five other matched control establishments. The standard errors are clustered by establishment ID for all our analyses.
Importantly, these types of variations in the model are due to varying assumptions about the underlying data generation process. In the results section, we discuss the varying combinations and how they impact our interpretation of the data.