Previous research has demonstrated the existence of a Flynn effect in the National Longitudinal Survey of Youth Children (NLSYC) responses to the PIAT-Math instrument. The PIAT-Math is at least partially linked to fluid intelligence, whereas other scales in the NLSYC - PIAT-Reading Recognition, PIAT-Reading Comprehension, the PPVT, and Digit Span - are primarily based on crystallized intelligence; these scales showed little or no Flynn effect in the NLSYC. We put the 84 PIAT-Math items "under the microscope" by evaluating the Flynn effect in relation to each item, and measured the Flynn effect by computing a slope across birth-year cohorts, using nine different age replications. Following, we use expert ratings of the items on eight different features - visual matching, recall/memory, computation/estimation, spatial visualization, real-world reasoning, manipulation of geometry, solving algebra, and counting - to identify what features are important in producing the Flynn effect. The highest correlations obtain for the links between the Flynn effect and the features real-world reasoning, counting and computation/estimation. There is a negative correlation between item-level Flynn effects and the features manipulation of geometry, solving algebra, and recall/memory. These results support previous findings ephasizing the role that fluid intelligence plays in relation to the Flynn effect.