The primary sources of data covering a period of January 1, 2003 to December 31, 2007 were: (1) the enrollment history of clients participating in 24 NFP programs throughout Pennsylvania; (2) birth certificate files from the Pennsylvania Department of Public Health; (3) death certificate files from the Pennsylvania Department of Public Health; and (4) welfare eligibility files from the Department of Public Welfare.
The target population was clients from the 24 NFP sites in Pennsylvania who were enrolled in the NFP program between January 1, 2003 and December 31, 2007. Included were women who: (1) delivered a first-born infant; (2) self-reported tobacco cigarette use in the first trimester of pregnancy on birth certificate; and (3) received welfare assistance from the Commonwealth of Pennsylvania within 12 months prior to the birth of their first-born infant.
Eligible women for an unexposed comparison group were identified following a previously described linkage to birth certificate and welfare eligibility data from women residing in NFP communities who met the inclusion criteria noted above
]. The unexposed comparison group identified women eligible for enrollment in NFP who were non-enrollees. Among comparison women, reasons for non-enrollment in NFP were not determinable at an individual-level, but likely include: program caps on yearly enrollment, receipt of alternate community services, and lack of interest or ability to enroll. To identify a comparison group from among the unexposed eligible women, a propensity score analysis used data from birth certificates and welfare eligibility files to model factors associated with a woman’s participation in NFP. The factors included maternal education (<12th, high school, some college or higher), maternal race (White, Black, Hispanic, Other), marital status (y/n), TANF receipt prior and/or during first trimester of pregnancy (y/n), foodstamp receipt prior and/or during first trimester of pregnancy (y/n), and history of gestational diabetes or hypertension (y/n). In addition, variables were included that encoded high density zipcodes within each site catchment area in order to drive the selection of unexposed comparison women toward high-penetration neighborhoods of interest. This density variable was created by identifying zipcodes as high density that enrolled more than 5% of the NFP client population within the site catchment area. Finally, models were stratified on maternal age (≤18 years, >18 years) and time period of birth cohort (2003–2005, 2006–2007) to force balancing on these factors for subsequent stratified analyses.
We developed a separate propensity score for each site to allow for the maximum flexibility in modeling of potential confounders with the fewest assumptions about the association of covariates and NFP participation across sites. Using a separate logistic regression for each site, the expected probability of participation in NFP (propensity score) was then estimated based on the above characteristics for each woman within a site
]. The next step excluded as potential matches all unexposed comparison women who had propensity scores that fell outside the range of propensity scores of NFP clients. This initial exclusion left a group of unexposed comparison women and a group of NFP clients who shared propensity scores with “common support” – or overlapping ranges of propensity scores. Using a caliper of 0.05 (probability scale), one or more unexposed comparison women were selected using a nearest neighbor match without resampling (up to a maximum of 4 matched comparison women per client). Matching was done with the program gmatch under the SAS® Statistical Package v9.1.3
]. To avoid bias in differential matching rates across clients, analysis weights were assigned to comparison women based on the number of unexposed women matched to the NFP client
]. Overall, successful balance of covariates was achieved within each site propensity score model. There were several exceptions, principally stemming from small sample size in which any single woman was able to shift weighted balance significantly. At a site model level, a threshold of ≥5% weighted case-comparison difference was used to identify potential covariate imbalance. Two agencies showed imbalance in the case-comparison weighted difference for race (10% and 13%), eight for education (range: 5-10%), five for prenatal foodstamp receipt (range: 5-9%), six for prenatal TANF receipt (range: 5-14%), and six for prenatal hypertension (range: 5-10%). In aggregate among all agencies, the study cohort was balanced across all covariates (see Table
Characteristics of nurse-family partnership clients compared with all potential comparison women, welfare-eligible comparison women, and final matched comparison women across the commonwealth of Pennsylvania
The primary outcome was a binary measure of smoking cessation in the third trimester of pregnancy. Smoking cessation was defined as self-reported use of zero cigarettes during the last three months of pregnancy as recorded on birth certificates. The two independent variables of interest were the assignment to the NFP program (yes/no) and a measure of the baseline county smoking rate where the site resided. Baseline county smoking rate was created for each client and was defined as the proportion of all women delivering in the county with birth certificate self-reported use of cigarettes in the first three months of pregnancy or in the three months prior to pregnancy where first trimester smoking information was missing (n
Multivariable logistic regression of the matched sample examined the association between smoking cessation and NFP participation. We hypothesized that smoking cessation success might also depend upon the prevalence of prenatal smoking in a woman’s community. Because prenatal smoking rates at the county level varied within site catchment area, we decomposed this factor into the site-level mean (to measure the across-site component) and the difference between mother’s zipcode prenatal smoking rate (calculated by estimating prevalence within zipcode) and the county rate (to measure the within site component)
]. We estimated and report robust variance estimates to account for the potential of lack of model fit owing to possible overdispersion of the data. Results were expressed as odds ratios and predictive margins
Analyses were conducted using Stata versions 11.0 (College Station, TX) and SAS v9.2 (SAS Institute, Cary NC). Approval for the study was granted by the Department of Public Welfare for the Commonwealth of Pennsylvania. Ethical approval was granted by the Institutional Review Board at the Children’s Hospital of Philadelphia.