Magruder (2003) used correlation analysis to study sales of two categories of OTC products: medications used to treat "flu'' symptoms (which they called "flu remedies''), and chest rubs. Similar to Campbell and colleagues, Magruder used billing data from physician office visits as a gold standard. He studied the National Capital Area that includes Washington DC and its surrounding suburbs for a 19-month period. He found that the peak correlation of chest rubs with physician office visits was 0.86, and that the correlation was maximal when sales of chest rubs preceded office visits by seven days. Similarly, the peak correlation of flu remedies with physician office visits was 0.89 when flu remedies preceded visits by 12 days. He was unable to compute a confidence interval on these measurements because he computed a single correlation for each OTC category for the entire National Capital Area. After accounting for holiday effects in the physician visit data and dividing the National Capital Area into six subregions, he found that the peak correlation of flu remedies with office visits averaged 0.895 (95% CI, 0.87-0.92) over the six sub-regions, and flu remedies preceded office visits by 2.8 days (95% CI, 0.09-5.58).
Das et al. (2005) used correlation analysis and the detection algorithm method to study OTC products for the treatment of influenza-like illness or ILI in New York City. They found that the ratio of sales of ILI products (a synonym for flu remedies) to sales of pain-relief products (the OTC ratio) was highly correlated with the ratio of ED visits for ILI to ED visits for other syndromes (the ED ratio), with an r2 of 0.60. This correlation occurred when neither time series was lagged relative to the other one. They also measured the correlation when OTC sales were lagged by -14, -7, 7, and 14 days, and found lower correlations at these lags, and thus a lag of zero produced the highest correlation (of the lags they studied). In their detection algorithm analysis, they were able to detect the 2003-2004 influenza outbreak approximately two to three weeks earlier from the ED ratio than from the OTC ratio using a cyclical regression model. Similarly, they were able to detect the 2004-2005 influenza outbreak approximately six days earlier from the ED ratio.
Of historical interest, a study by Welliver et al. (1979) was the first to provide a quantitative estimate of the lead time of OTC sales data over data collected during physician office visits. Welliver and colleagues observed a strong peak in sales of cold remedies (another synonym for flu remedies) just before the rise in physician encounters with patients subsequently diagnosed as having influenza B virus, and one week before the peak in those encounters. They also observed an earlier rise in sales of cold remedies approximately coincident with the early winter rise in non-influenza respiratory virus activity as evidenced by laboratory test results. They found no correlation between sales of antifever medications, such as aspirin (either adult or pediatric), and influenza. Welliver's study was based on two outbreaks in a single city, and utilized OTC data that were aggregated at the weekly level.
Was this article helpful?