Changes in Cancer Screening Rates Following a New Cancer Diagnosis in a Primary Care Patient Panel

Key Points Question Is primary care physician (PCP) exposure to a patient with a new breast or colorectal cancer diagnosis associated with increases in cancer screenings for other patients who subsequently visit the affected PCP? Findings In this cohort study of 3158 PCPs caring for 1 920 189 patients, using stacked difference-in-differences analyses, there were rapid and sustained increases in cancer screening rates for patients visiting PCPs who were recently exposed to new cancer diagnoses, for both breast cancer and colorectal cancer. Meaning These findings suggest that PCPs’ exposures to new diagnoses of cancer are associated with significant, sustained increases in cancer screening rates for other patients subsequently visiting the exposed PCPs.

To construct the analytical dataset, for a given quarter, we denote PCPs who experience an exposure in the current quarter as treatment PCPs, and PCPs who experience an exposure in the future (more than 4 quarters in the future) as comparison PCPs (eFigure). The sample was constructed by repeating this procedure for each of the 84 quarters over the entire data period, then appending the resulting quarter-specific datasets into one dataset. Note that PCPs can only be in the treatment group in one quarter (PCP's quarter of earliest exposure to a new cancer diagnosis) but can be in the comparison group in multiple quarters. Thus, the sum of currently-exposed PCPs and future-exposed PCPs (eTable 4) is greater than the number of individual PCPs in the study sample.
To exclude outlier physicians with few attributed patients, physicians in the lowest fifth percentile of patient panel size were excluded. In the breast cancer analysis, the mean annual patient panel size for PCPs in the lowest fifth percentile of panel size was 38.7 patients (SD = 4.7). In the colorectal cancer analysis, the mean annual patient panel size for PCPs in the lowest fifth percentile of panel size was 54.4 patients (SD = 10.8).
The outcome measure was calculated at the PCP-quarter level, representing the proportion of patients who visited a PCP in a given quarter who received a relevant cancer screening within a year of that PCP visit. The stacked difference-in-differences model calculated the difference between treatment and comparison PCPs in screening rates (outcome measure) before vs. after exposure to a new cancer diagnosis. The quarter of exposure (relative quarter 0) was included in the exposure period because it is possible for some effect to already occur during the quarter of exposure. The model included a vector of calendar month indicator variables (calendar month fixed effects), year indicator variables (year fixed effects), and PCP indicator variables (PCP fixed effects).
Additionally, we estimated a "relative quarter" version of the main model, where each relative quarter coefficient represented the difference between treatment and comparison PCPs in screening rates in that quarter relative to the quarter prior to exposure to a new cancer diagnosis (relative quarter -1).
Subgroup analyses, which estimated the relative effect sizes of subgroups of PCP characteristics, were estimated through the coefficient of an interaction term between the subgroup characteristic of interest (as an indicator variable) and the main model effect coefficient.
All models were estimated using robust standard errors clustered at the PCP level.

F. Tests of Pre-Exposure Parallel Trends
An identifying assumption for difference-in-differences requires that in the absence of study exposure, differences in outcomes for the treatment and comparison PCP groups would remain constant over time. Though this assumption cannot be explicitly tested, a test of pre-exposure parallel trends can aid in examining the likelihood that it holds. We tested for differential changes between treatment and comparison groups prior to exposure, using F-tests of preexposure coefficients, where a null finding provides support for parallel trends (by signifying there is no evidence to reject the null of parallel trends). Null findings demonstrate that the differences in outcomes for these groups were not showing significant signs of deviating from each other for a considerable period before exposure (-Q4 through -Q1). Specifically, the test assessed whether the difference-in-difference between relative quarters -Q4, -Q3, -Q2, and -Q1 (pre-exposure, i.e., before Q0) was different for PCPs who would be treated at Q0 (treatment, currently-exposed PCPs) vs. PCPs who would be treated later than Q0 (comparison, futureexposed PCPs). As noted previously, Q0 was included in the exposure period because it is possible for some effect to already occur during the quarter of exposure.

F. Falsification Tests for Breast Cancer and Colorectal Cancer Analyses
In the breast cancer falsification test, colorectal cancer diagnoses were used as index events for breast cancer screening rates in the stacked DD model as specified in the main breast cancer analysis. In the colorectal cancer falsification test, breast cancer diagnoses were used as index events for colorectal cancer screening rates in the stacked DD model as specified in the main colorectal cancer analysis. Results are included in eTable 5.

G. Sensitivity Analyses and Additional Analyses
We conducted additional analyses using varying specifications, including sample and exposure criteria, as detailed below.  b N=898 unique PCPs in breast cancer analysis. N=370 unique PCPs in colorectal cancer analysis. PCPs in the stacked DD model can act as both currently-exposed and future-exposed PCPs at different time points. Thus, the column sums of currently-and futureexposed PCPs in the table are greater than the total number of individual exposed PCPs.    a Colorectal cancer analyses were stratified into three cancer modality categories: 1) "Colonoscopy:" colonoscopy only; 2) "FOBT/FIT:" fecal occult blood testing (FOBT), fecal immunochemical testing (FIT), and/or multitarget stool DNA testing; and 3) "Sigmoidoscopy:" sigmoidoscopy only. Colonoscopy analysis: P = 0.12 for joint significance test of pre-exposure coefficients. FOBT/FIT analysis: P = 0.68 for joint significance test of pre-exposure coefficients. Sigmoidoscopy analysis: P = 0.51 for joint significance test of pre-exposure coefficients.