We first focused on simple comparisons of trends in gonorrhea incidence rates in 2 groups of cities in the United States: those with relatively high prevalence and those with relatively low prevalence of ciprofloxacin resistance. After performing these illustrative comparisons, we used regression analyses to examine the association between ciprofloxacin resistance and gonorrhea incidence in a more robust manner. For simplicity, we describe our study as a city-level analysis, although as described in more detail below, the data we analyzed comprised a mixture of sources at the city, county, and metropolitan statistical area levels.

We used antimicrobial drug susceptibility data from the Gonococcal Isolate Surveillance Project (GISP) to analyze the association between ciprofloxacin resistance and gonorrhea incidence over time at the city level. GISP has been described in detail by Schwarcz et al. (11) and Kirkcaldy et al. (12). In brief, GISP is a sentinel surveillance system established in 1986 to monitor antimicrobial drug susceptibility among N. gonorrhoeae isolates. Each month, urethral gonococcal isolates and clinical data are systematically collected consecutively from up to the first 25 men at participating STD clinics in each city in whom urethral gonorrhea was diagnosed (11,12). Cities may have a single participating clinic or multiple clinics. The gonococcal isolates are tested for antimicrobial drug susceptibility by using agar dilution method. Ciprofloxacin susceptibility has been monitored in GISP since 1990 (9). The prevalence of ciprofloxacin resistance increased during the late 1990s and 2000s (9). By 2007, ciprofloxacin resistance was prevalent in all regions of the United States, prompting CDC to no longer recommend fluoroquinolones for treatment of gonorrhea (9). Given this time line and availability of data for additional variables described later, we included the years 1991–2006 in our analysis.

Cities with ≥1 STD clinic participating in GISP were included in the study if annual ciprofloxacin resistance prevalence data were available from that city for ≥13 years during 1991–2006 and if city-level gonorrhea incidence rates during the same 16-year period were available from gonorrhea case report surveillance data maintained by the Division of STD Prevention at CDC. Seventeen cities met the inclusion criteria: Albuquerque (New Mexico), Atlanta (Georgia), Baltimore (Maryland), Birmingham (Alabama), Cincinnati (Ohio), Cleveland (Ohio), Denver (Colorado), Honolulu (Hawaii), Minneapolis (Minnesota), New Orleans (Louisiana), Philadelphia (Pennsylvania), Phoenix (Arizona), Portland (Oregon), San Diego (California), San Francisco (California), Seattle (Washington), and St. Louis (Missouri). For most cities in our analysis, the city-specific STD rates we obtained were derived from county data and may only approximate city jurisdictions. Our dataset consisted of 272 observations, and each observation included the annual prevalence of city-level gonococcal ciprofloxacin resistance (prevalence in 17 cities each year over a 16-year period).

We calculated the median percentage of isolates resistant to ciprofloxacin in 2004 and labeled the 8 cities above the median as higher resistance cities and the 9 cities at or below the median as lower resistance cities. For each group, we calculated gonorrhea incidence rates during 1991–2006. The rate for each group of cities was calculated as the sum of reported gonorrhea cases in the cities divided by the sum of the populations of the cities and multiplied by 100,000. The percentage of isolates resistant to ciprofloxacin for each group of cities was calculated as the average across all cities in the group.

We performed regression analyses in which the dependent variable was the city gonorrhea incidence rate (log) and the independent variable of interest was the percentage of GISP isolates resistant to ciprofloxacin in GISP clinic(s) located in the given city. The regression also included sociodemographic variables (percentage of persons who were black, percentage of persons 15–29 years of age, unemployment rate, per capita income, robbery rate) and binary (dummy) variables for each city and year to control for city-specific factors and national trends in factors that influence city-level gonorrhea incidence rates (Table 1).

We included percentage of persons who were black and percentage of persons 15–29 years of age as explanatory variables because reported STD rates are often disproportionately high among black persons and youth (13). We included unemployment, income, and robbery rates as explanatory variables because STD rates also have been linked to social determinants of health (14,15). Sociodemographic variables, such as these, have been shown to correlate with STD rates at the population level over time (16,17).

Gonorrhea and syphilis incidence rates, percentage of persons who were black, and percentage of persons 15–29 years of age were obtained from surveillance records and US Census Bureau data maintained by CDC (13). Robbery rates, unemployment rates, and per capita income data were obtained online from various federal agencies (Table 1). Of our 272 observations, 11 had missing values for ≥1 variable. Missing values for the variables were replaced with estimated values, and we assumed a linear trend from one year to the next. For example, if the unemployment rate for a given city in 2004 was missing, the average of the unemployment rate for the given city in 2003 and 2005 was assigned for 2004.

A common problem with regression analysis of data consisting of multiple observations over time is serial correlation, in which the error term in a given year correlates with the error term in the previous year. We used 2 approaches to address the issue of serial correlation. First, we calculated SEs that are robust to the serial correlation. Second, we corrected for the autocorrelated error terms when computing the regression (18). Specifically, we used ordinary least squares (OLS), included the lagged dependent variable as an exploratory variable, and used the Newey-West procedure to calculate heteroskedasticity- and autocorrelation-consistent SEs for the regression coefficients. We also estimated a linear regression with correction for first-order autocorrelated errors (AR1) by using the AR1 procedure.

The specific equation we estimated with OLS was G_i,t = α + β₁G_i,t-1 + β₂R_i,t + γX_i,t + C_i + Y_t + ε_i,t, in which G_i,t is the log of the gonorrhea incidence rate in city i in year t, α is a constant, R_i,t is the percentage of isolates resistant to ciprofloxacin in GISP clinic(s) of city i in year t, X_i,t is a vector of sociodemographic variables listed earlier, C denotes city dummy variables, Y denotes year dummy variables, and ε is the error term. The equation we estimated with AR1 was the same as the previous equation except that the lagged value of the dependent variable (G_{i,t – 1}) was not included in the model. Thus, the differences between the 2 approaches we used to address serial correlation can be summarized as follows. The OLS regression includes the lagged value of gonorrhea incidence rates as an independent variable and calculates SEs that are robust to autocorrelation in the error terms. The AR1 regression is corrected for first-order correlation in the error terms and does not include the lagged value of the gonorrhea incidence rate. Analyses were conducted by using WinRATS version 8.01 (Estima, Evanston, IL, USA).

We performed additional analyses to examine the robustness of our results. First, we repeated our regression analysis by substituting the log of the syphilis rate for the log of the gonorrhea incidence rate as the dependent variable, thereby testing to determine whether our model would suggest an implausible link between gonococcal ciprofloxacin resistance and changes in the incidence of syphilis. In performing this procedure, we added 1 to the syphilis rate before taking the log so as not to exclude observations in which the syphilis rate was 0. Second, we examined temporal aspects of the association between ciprofloxacin resistance and gonorrhea incidence rates to determine whether gonorrhea incidence rates could be better predicted on the basis of past values of gonorrhea incidence rates and ciprofloxacin resistance rather than past values of gonorrhea incidence rates alone (as in Granger causality tests) (18,19). To do so, we modified our model so that 3 lagged values of the resistance variable (R_{i,t – 1}, R_{i,t –2}, and R_{i,t – 3}) were included as explanatory variables rather than the current year value of the resistance variable (R_i,t). We also included 3 lagged values of gonorrhea incidence (specifically, the log of the gonorrhea incidence rate in years t – 3, t –2, and t –1) as explanatory variables rather than 1 lag. We examined the joint significance of the 3 lagged resistance variables (R_{i,t – 1}, R_{i,t – 2}, and R_{i,t – 3}) by using an F test to compare this model with a restricted model in which the coefficients of these 3 variables were set to 0. The joint significance of the 3 lagged values of gonorrhea incidence was calculated in an analogous manner. We then reversed the model such that ciprofloxacin resistance was the dependent variable. Third, we tested the sensitivity of our results to functional form by using the gonorrhea incidence rate rather than the log of the gonorrhea incidence rate as the dependent variable. Fourth, we tested for the effect of influential observations by using 2 approaches: deleting observations with a residual >2 SEs and repeating the main analysis 17 times, each time omitting 1 of the 17 cities from the analysis.

In 2004, a median percentage of 3.3% of isolates were resistant to ciprofloxacin in the 17 cities in our analysis. We classified the 8 cities above the median in 2004 as higher resistance cities and the 9 cities at or below the median in 2004 as lower resistance cities. Cities with higher resistance were Denver, Honolulu, Minneapolis, Phoenix, Portland, San Diego, San Francisco, and Seattle. Cities with lower resistance were Albuquerque, Atlanta, Baltimore, Birmingham, Cincinnati, Cleveland, New Orleans, Philadelphia, and St. Louis. In our simple comparison of higher resistance and lower resistance cities, we found divergent trends in gonorrhea incidence rates in the 2000s (Figure). Although gonorrhea incidence rates were much lower overall in the higher resistance cities, gonorrhea incidence rates generally increased in the higher resistance cities and decreased in the lower resistance cities during 2000–2006 (Figure, panel A). The timing of the divergent trends in gonorrhea incidence rates coincided with the divergent trends in ciprofloxacin resistance (Figure, panel B).

The coefficient of the ciprofloxacin resistance variable was positive and significant across all 4 models we estimated (p<0.01) (Table 2). Ciprofloxacin resistance in a given city in a given year was associated with higher gonorrhea incidence rates in that city in the given year. This finding was consistent regardless of estimation procedure (OLS in models 1 and 2 and AR1 in models 3 and 4) and regardless of the exclusion (models 1 and 3) or inclusion (models 2 and 4) of the additional sociodemographic variables.

We found no association between ciprofloxacin resistance and syphilis incidence. When we examined the temporal association between ciprofloxacin resistance and gonorrhea incidence, the coefficients of the lagged individual ciprofloxacin resistance variables were not all significant individually when the dependent variable was the log of the gonorrhea incidence rate (Table 3). However, the sum of the coefficients of the lagged ciprofloxacin resistance variables was positive and these coefficients were jointly significant (p<0.01). When we reversed the model such that ciprofloxacin resistance was the dependent variable, lagged values of the gonorrhea incidence coefficients (specifically the coefficients of the logs of the gonorrhea incidence rate in years t – 3, t – 2, and t – 1) were not jointly significant (Table 3). Although past levels of ciprofloxacin resistance helped to predict current gonorrhea incidence rates, past gonorrhea incidence rates did not help to predict current ciprofloxacin resistance levels. Our results were generally consistent across the range of additional analyses we conducted, including applying gonorrhea incidence rates in non-log form, omitting outliers, and omitting any given city from the analysis.