IP fragmentation can cause excessive retransmission at the TCP level.
TCP transmits information as a series of segments, and these are the units of acknowledgement and retransmission as well. If a TCP segment is lost in the network, the entire segment has to be retransmitted.
If IP fragmentation occurs, the segment will be split into multiple fragments. The receiving system can't acknowledge a segment until it receives all its fragments.
This means that if a single fragment is lost, the entire segment that it was a part of will need to be retransmitted. To avoid this potential problem, TCP tries to size its segments so they will not need to be fragmented. That way, if a packet is lost, only that packet will have to be retransmitted.
Retransmission generally has a dramatic impact on TCP throughput. The design of TCP assumes that packet loss is usually due to network congestion, and the way to avoid this is to reduce the amount of network bandwidth being used, i.e. slow down transmission. So when segments are lost, we have to resend them (using more network resources) and we slow down. These are both detrimental, so we try to minimize the amount of retransmission.
Making the segment size equal to the Path MTU (minus the TCP and IP header lengths) makes it less likely that a segment will need to be retransmitted. Imagine that 1 out of 100 packets is lost. If each packet is just one segment, it means we only retransmit 1% of segments. But if fragmentation is causing segments to be split into an average of 5 packets, we'll retransmit 5% of segments.